{
  "ctfidf_model": {
    "bm25_weighting": false,
    "reduce_frequent_words": false
  },
  "vectorizer_model": {
    "params": {
      "analyzer": "word",
      "binary": false,
      "decode_error": "strict",
      "encoding": "utf-8",
      "input": "content",
      "lowercase": true,
      "max_df": 1.0,
      "max_features": null,
      "min_df": 2,
      "ngram_range": [
        1,
        5
      ],
      "stop_words": "english",
      "strip_accents": null,
      "token_pattern": "(?u)\\b\\w\\w+\\b",
      "vocabulary": null
    },
    "vocab": {
      "making": 19496,
      "bert": 3494,
      "neural": 22719,
      "machine": 19339,
      "translation": 33817,
      "gpt2": 12866,
      "demonstrate": 7432,
      "effectiveness": 8937,
      "using": 34720,
      "pretrained": 25625,
      "language": 16032,
      "models": 20917,
      "lms": 19066,
      "various": 35072,
      "natural": 22502,
      "processing": 26093,
      "tasks": 32225,
      "lm": 19054,
      "finetuning": 11366,
      "suffers": 31557,
      "catastrophic": 4212,
      "forgetting": 11730,
      "applied": 2185,
      "resourcerich": 28427,
      "work": 35662,
      "introduce": 15496,
      "training": 33437,
      "framework": 11823,
      "key": 15752,
      "integrate": 15319,
      "nmt": 22977,
      "proposed": 26587,
      "consists": 5761,
      "techniques": 32625,
      "distillation": 8334,
      "ensure": 9600,
      "model": 20330,
      "retain": 28716,
      "previous": 25864,
      "knowledge": 15808,
      "dynamic": 8756,
      "switching": 31798,
      "avoid": 3036,
      "strategy": 31113,
      "adjust": 1389,
      "learning": 17529,
      "according": 845,
      "scheduled": 29227,
      "policy": 25079,
      "experiments": 10415,
      "gains": 12069,
      "bleu": 3764,
      "score": 29264,
      "wmt14": 35630,
      "englishgerman": 9497,
      "pair": 23969,
      "surpasses": 31737,
      "stateoftheart": 30919,
      "pretraining": 25783,
      "aided": 1627,
      "14": 134,
      "large": 16926,
      "task": 32069,
      "40": 373,
      "millions": 20167,
      "base": 3113,
      "significantly": 30029,
      "improves": 14368,
      "transformer": 33701,
      "big": 3698,
      "code": 4720,
      "downloaded": 8670,
      "release": 27898,
      "strategies": 31100,
      "social": 30416,
      "impacts": 14146,
      "range": 27185,
      "beneficial": 3478,
      "uses": 34707,
      "assist": 2632,
      "prose": 26652,
      "poetry": 25061,
      "programming": 26198,
      "analyze": 1987,
      "dataset": 6932,
      "biases": 3671,
      "flexibility": 11615,
      "generative": 12644,
      "capabilities": 3999,
      "raise": 27162,
      "misuse": 20244,
      "concerns": 5540,
      "report": 28111,
      "discusses": 8297,
      "openais": 23437,
      "related": 27853,
      "allows": 1809,
      "time": 33107,
      "releases": 27929,
      "conduct": 5581,
      "risk": 28892,
      "benefit": 3479,
      "analyses": 1907,
      "sizes": 30292,
      "increased": 14608,
      "ongoing": 23355,
      "research": 28285,
      "provides": 26746,
      "recommendations": 27652,
      "better": 3586,
      "coordination": 6138,
      "responsible": 28519,
      "publication": 26845,
      "ai": 1593,
      "masked": 19609,
      "scoring": 29285,
      "mlms": 20302,
      "require": 28209,
      "nlp": 22918,
      "instead": 15119,
      "evaluate": 9816,
      "box": 3847,
      "scores": 29277,
      "computed": 5499,
      "masking": 19619,
      "tokens": 33215,
      "outperform": 23761,
      "autoregressive": 2935,
      "like": 17843,
      "variety": 35056,
      "rescoring": 28283,
      "asr": 2582,
      "hypotheses": 13962,
      "roberta": 28914,
      "reduces": 27731,
      "endtoend": 9432,
      "librispeech": 17813,
      "wer": 35534,
      "30": 305,
      "relative": 27876,
      "adds": 1376,
      "17": 169,
      "baselines": 3261,
      "lowresource": 19310,
      "pairs": 23974,
      "domain": 8550,
      "adaptation": 1173,
      "attribute": 2756,
      "success": 31507,
      "unsupervised": 34453,
      "expression": 10642,
      "linguistic": 18003,
      "acceptability": 810,
      "lefttoright": 17693,
      "bias": 3641,
      "greatly": 13271,
      "improving": 14402,
      "10": 24,
      "points": 25069,
      "effects": 8977,
      "licensing": 17817,
      "blimp": 3771,
      "finetune": 11280,
      "enabling": 9315,
      "computation": 5442,
      "single": 30195,
      "inference": 14758,
      "pass": 24371,
      "associated": 2642,
      "enable": 9283,
      "plugandplay": 25055,
      "use": 34512,
      "growing": 13310,
      "number": 23138,
      "crosslingual": 6412,
      "translations": 33858,
      "multiple": 22376,
      "languages": 16861,
      "library": 17812,
      "commongen": 5022,
      "constrained": 5779,
      "text": 32812,
      "generation": 12448,
      "challenge": 4307,
      "commonsense": 5034,
      "reasoning": 27367,
      "recently": 27585,
      "largescale": 17342,
      "demonstrated": 7514,
      "impressive": 14228,
      "performance": 24511,
      "benchmark": 3349,
      "datasets": 7058,
      "building": 3918,
      "machines": 19373,
      "compose": 5319,
      "plausible": 25014,
      "sentences": 29549,
      "remains": 27991,
      "challenging": 4378,
      "paper": 24001,
      "present": 25510,
      "explicitly": 10547,
      "test": 32756,
      "ability": 650,
      "given": 12740,
      "set": 29669,
      "common": 5002,
      "concepts": 5530,
      "dog": 8545,
      "catch": 4218,
      "throw": 33100,
      "generate": 12260,
      "coherent": 4890,
      "sentence": 29527,
      "describing": 7681,
      "everyday": 10054,
      "scenario": 29198,
      "inherently": 14953,
      "requires": 28244,
      "relational": 27867,
      "background": 3062,
      "compositional": 5326,
      "generalization": 12203,
      "unseen": 34432,
      "concept": 5526,
      "combinations": 4953,
      "constructed": 5805,
      "combination": 4948,
      "crowdsourced": 6431,
      "existing": 10260,
      "caption": 4133,
      "corpora": 6163,
      "descriptions": 7685,
      "unique": 34354,
      "gap": 12086,
      "t5": 31934,
      "human": 13789,
      "furthermore": 11984,
      "learned": 17518,
      "capability": 4086,
      "transferred": 33689,
      "improve": 14252,
      "downstream": 8671,
      "generating": 12407,
      "additional": 1249,
      "context": 5879,
      "entities": 9629,
      "understanding": 34205,
      "recent": 27486,
      "progress": 26206,
      "witnessed": 35624,
      "development": 7948,
      "gpt": 12840,
      "xlnet": 35878,
      "based": 3130,
      "et": 9788,
      "al": 1679,
      "2017": 220,
      "end": 9408,
      "achieved": 998,
      "results": 28565,
      "approaching": 2404,
      "demonstrates": 7555,
      "power": 25318,
      "selfattention": 29407,
      "architecture": 2435,
      "paired": 23972,
      "sufficient": 31559,
      "layers": 17436,
      "data": 6585,
      "complex": 5265,
      "cues": 6456,
      "2018": 221,
      "showed": 29855,
      "possible": 25211,
      "inject": 14978,
      "syntactic": 31813,
      "structure": 31212,
      "supervised": 31670,
      "conjecture": 5676,
      "similar": 30095,
      "injection": 14982,
      "semantic": 29445,
      "particular": 24334,
      "coreference": 6155,
      "information": 14852,
      "problems": 26022,
      "lambada": 16019,
      "2016": 219,
      "trained": 33382,
      "scratch": 29289,
      "auxiliary": 2959,
      "supervision": 31695,
      "outperforms": 23804,
      "largest": 17389,
      "setting": 29721,
      "new": 22772,
      "containing": 5833,
      "tiny": 33171,
      "fraction": 11815,
      "parameters": 24219,
      "compared": 5117,
      "thorough": 33070,
      "analysis": 1910,
      "different": 8042,
      "variants": 35041,
      "architectures": 2459,
      "configurations": 5656,
      "suggesting": 31582,
      "future": 12027,
      "directions": 8227,
      "applying": 2216,
      "gorilla": 12834,
      "camel": 3987,
      "plausibility": 25013,
      "modeling": 20885,
      "world": 35832,
      "used": 34582,
      "testbed": 32793,
      "exploring": 10614,
      "representations": 28156,
      "focused": 11660,
      "specifically": 30726,
      "physical": 24940,
      "shown": 29866,
      "distributional": 8399,
      "methods": 19998,
      "fail": 10898,
      "tested": 32794,
      "led": 17682,
      "improved": 14304,
      "fact": 10855,
      "effective": 8858,
      "difficult": 8168,
      "problem": 25985,
      "directly": 8232,
      "create": 6346,
      "extracting": 10756,
      "events": 10047,
      "corpus": 6173,
      "provide": 26681,
      "baseline": 3239,
      "selfsupervised": 29436,
      "manner": 19546,
      "testing": 32799,
      "believe": 3338,
      "injecting": 14981,
      "explicit": 10544,
      "bloom": 3781,
      "filter": 11170,
      "meets": 19784,
      "extend": 10646,
      "idea": 13984,
      "word": 35633,
      "pieces": 24950,
      "opaque": 23380,
      "ids": 14034,
      "hash": 13470,
      "functions": 11970,
      "map": 19576,
      "id": 13981,
      "smaller": 30373,
      "space": 30578,
      "similarly": 30133,
      "multilayer": 22292,
      "able": 740,
      "obtain": 23245,
      "high": 13548,
      "accuracy": 857,
      "size": 30240,
      "degree": 7386,
      "larger": 17314,
      "sampled": 29067,
      "softmax": 30451,
      "computational": 5452,
      "budget": 3901,
      "observation": 23217,
      "important": 14195,
      "remove": 28074,
      "ambiguity": 1873,
      "input": 15002,
      "alternative": 1848,
      "method": 19866,
      "solving": 30505,
      "vocabulary": 35381,
      "observed": 23235,
      "unstructured": 34450,
      "implicitly": 14178,
      "store": 31079,
      "retrieve": 28766,
      "queries": 27018,
      "short": 29809,
      "measure": 19732,
      "practical": 25359,
      "utility": 34951,
      "approach": 2224,
      "answer": 2039,
      "questions": 27095,
      "access": 815,
      "external": 10726,
      "scales": 29150,
      "performs": 24840,
      "competitively": 5232,
      "opendomain": 23449,
      "systems": 31886,
      "answers": 2082,
      "source": 30546,
      "answering": 2055,
      "facilitate": 10834,
      "reproducibility": 28202,
      "augmentation": 2795,
      "provided": 26736,
      "significant": 29949,
      "study": 31293,
      "types": 34057,
      "autoencoder": 2842,
      "seq2seq": 29588,
      "bart": 3106,
      "conditional": 5567,
      "prepending": 25500,
      "class": 4581,
      "labels": 15966,
      "sequences": 29612,
      "simple": 30139,
      "way": 35428,
      "condition": 5566,
      "additionally": 1268,
      "classification": 4590,
      "benchmarks": 3428,
      "explore": 10570,
      "differs": 8167,
      "diversity": 8476,
      "preserve": 25602,
      "cooking": 6131,
      "recipe": 27629,
      "evaluation": 9915,
      "interests": 15414,
      "automatic": 2875,
      "recipes": 27632,
      "past": 24387,
      "years": 35889,
      "thanks": 33043,
      "online": 23359,
      "novel": 23051,
      "modes": 22175,
      "generations": 12641,
      "instruction": 15141,
      "title": 33177,
      "ingredients": 14944,
      "ingredient": 14943,
      "instructions": 15241,
      "backend": 3061,
      "module": 22192,
      "comprises": 5431,
      "finetuned": 11303,
      "users": 34683,
      "quality": 26938,
      "generated": 12340,
      "contents": 5878,
      "reference": 27771,
      "accessed": 828,
      "trec": 33876,
      "cast": 4208,
      "2019": 222,
      "conversational": 6092,
      "assistance": 2634,
      "track": 33326,
      "overview": 23954,
      "seeking": 29359,
      "reusable": 28784,
      "collection": 4926,
      "search": 29306,
      "document": 8501,
      "passages": 24381,
      "retrieval": 28729,
      "car": 4157,
      "microsoft": 20157,
      "reading": 27304,
      "comprehension": 5340,
      "marco": 19581,
      "dialogues": 8023,
      "train": 33359,
      "50": 418,
      "average": 3007,
      "long": 19166,
      "relevance": 27932,
      "assessments": 2620,
      "topics": 33289,
      "20": 206,
      "year": 35888,
      "21": 249,
      "groups": 13303,
      "submitted": 31432,
      "total": 33297,
      "65": 474,
      "runs": 29015,
      "varying": 35167,
      "query": 27025,
      "ranking": 27229,
      "include": 14446,
      "traditional": 33341,
      "feature": 11020,
      "enhanced": 9531,
      "theme": 33045,
      "bertbased": 3540,
      "reranking": 28278,
      "leading": 17473,
      "employed": 9256,
      "expansion": 10344,
      "rewriting": 28867,
      "manually": 19561,
      "resolved": 28402,
      "utterances": 34979,
      "35": 339,
      "improvement": 14321,
      "manual": 19552,
      "rewrites": 28866,
      "best": 3553,
      "question": 27037,
      "reformulation": 27802,
      "sequencetosequence": 29618,
      "presents": 25572,
      "empirical": 9217,
      "plms": 25038,
      "leverage": 17742,
      "address": 1306,
      "strong": 31158,
      "independence": 14668,
      "assumption": 2663,
      "objective": 23201,
      "maximum": 19705,
      "likelihood": 17898,
      "estimation": 9786,
      "taskoriented": 32217,
      "dialogue": 8009,
      "indomain": 14723,
      "validate": 34991,
      "outdomain": 23739,
      "examining": 10107,
      "numbers": 23174,
      "texttotext": 33010,
      "transfer": 33667,
      "achieves": 1024,
      "fewer": 11084,
      "propose": 26489,
      "written": 35861,
      "situation": 30238,
      "real": 27309,
      "person": 24878,
      "currently": 6540,
      "facing": 10854,
      "helpful": 13516,
      "advice": 1528,
      "tests": 32807,
      "fundamental": 11971,
      "aspect": 2562,
      "resolve": 28401,
      "openended": 23456,
      "situations": 30239,
      "communicating": 5047,
      "todays": 33183,
      "struggle": 31235,
      "multibillion": 22269,
      "parameter": 24175,
      "examples": 10116,
      "writes": 35848,
      "humanwritten": 13936,
      "cases": 4197,
      "gpt3": 12979,
      "does": 8521,
      "worse": 35838,
      "low": 19265,
      "reveals": 28813,
      "errors": 9720,
      "hard": 13420,
      "spot": 30820,
      "outside": 23900,
      "showing": 29858,
      "room": 28966,
      "augmented": 2810,
      "relation": 27863,
      "extraction": 10761,
      "realworld": 27333,
      "deal": 7208,
      "limited": 17936,
      "imbalance": 14099,
      "issues": 15664,
      "augment": 2789,
      "properly": 26469,
      "specific": 30682,
      "gold": 12813,
      "classifier": 4623,
      "series": 29631,
      "advantages": 1503,
      "leads": 17488,
      "improvements": 14352,
      "11": 67,
      "f1": 10814,
      "state": 30902,
      "art": 2512,
      "widely": 35568,
      "biomedical": 3744,
      "surpassing": 31752,
      "47": 402,
      "geppetto": 12729,
      "italian": 15677,
      "available": 2964,
      "mainly": 19401,
      "english": 9474,
      "develop": 7907,
      "built": 3930,
      "means": 19729,
      "humanbased": 13889,
      "assessment": 2615,
      "calculating": 3964,
      "perplexity": 24872,
      "genres": 12722,
      "ii": 14041,
      "profiling": 26189,
      "writing": 35849,
      "characteristics": 4425,
      "production": 26170,
      "version": 35230,
      "shorter": 29825,
      "performed": 24826,
      "completion": 5258,
      "output": 23863,
      "judged": 15721,
      "closer": 4694,
      "original": 23698,
      "texts": 32988,
      "simpler": 30166,
      "datatotext": 7194,
      "pretrain": 25621,
      "indicate": 14686,
      "form": 11738,
      "enables": 9292,
      "pipelined": 24971,
      "tailored": 32015,
      "importantly": 14217,
      "evidenced": 10069,
      "outofdomain": 23750,
      "sets": 29717,
      "hope": 13753,
      "serves": 29653,
      "useful": 34638,
      "prevalent": 25856,
      "sense": 29505,
      "investigating": 15607,
      "adapterbased": 1203,
      "transformers": 33774,
      "following": 11685,
      "major": 19435,
      "structured": 31218,
      "resources": 28428,
      "hand": 13403,
      "joint": 15710,
      "adding": 1231,
      "objectives": 23209,
      "primary": 25914,
      "prohibitively": 26239,
      "computationally": 5484,
      "expensive": 10357,
      "posthoc": 25226,
      "lead": 17461,
      "investigate": 15570,
      "complementing": 5246,
      "conceptual": 5533,
      "conceptnet": 5529,
      "corresponding": 6224,
      "open": 23381,
      "mind": 20174,
      "respectively": 28454,
      "adapter": 1198,
      "overall": 23905,
      "glue": 12801,
      "picture": 24947,
      "deeper": 7347,
      "substantially": 31479,
      "type": 34052,
      "sourced": 30572,
      "team": 32598,
      "deep": 7316,
      "unifies": 34343,
      "attempt": 2702,
      "competition": 5216,
      "validation": 35004,
      "explanation": 10535,
      "discuss": 8289,
      "prepared": 25499,
      "labeled": 15954,
      "textual": 33021,
      "curated": 6468,
      "subtasks": 31497,
      "goal": 12805,
      "subtask": 31496,
      "distinguish": 8373,
      "statements": 30915,
      "make": 19450,
      "compare": 5103,
      "classifiers": 4627,
      "inspired": 15093,
      "questionanswering": 27081,
      "treat": 33871,
      "choice": 4550,
      "boost": 3814,
      "experimental": 10382,
      "second": 29319,
      "select": 29375,
      "reason": 27354,
      "statement": 30914,
      "stand": 30865,
      "teams": 32600,
      "937": 598,
      "27": 281,
      "participants": 24325,
      "competitive": 5217,
      "result": 28538,
      "nonsense": 23007,
      "shows": 29919,
      "potentials": 25317,
      "researches": 28381,
      "powerful": 25335,
      "fewshot": 11095,
      "aims": 1657,
      "reformulate": 27799,
      "concise": 5548,
      "fully": 11952,
      "specified": 30766,
      "effectively": 8908,
      "handled": 13411,
      "rules": 29009,
      "weak": 35455,
      "amounts": 1882,
      "ad": 1157,
      "hoc": 13732,
      "sessions": 29668,
      "rewrite": 28865,
      "weakly": 35460,
      "12": 87,
      "zeroshot": 35949,
      "gives": 12784,
      "comparable": 5072,
      "reveal": 28789,
      "syntax": 31831,
      "learns": 17677,
      "capture": 4143,
      "dependencies": 7616,
      "involve": 15624,
      "group": 13297,
      "references": 27773,
      "knowledgeaware": 15925,
      "hold": 13733,
      "adept": 1378,
      "semantics": 29487,
      "unclear": 34123,
      "grasp": 13243,
      "incorporate": 14568,
      "changing": 4413,
      "inserting": 15058,
      "storage": 31075,
      "simply": 30182,
      "signal": 29942,
      "existence": 10259,
      "tokenizer": 33211,
      "entity": 9643,
      "prediction": 25420,
      "solely": 30461,
      "signals": 29943,
      "packed": 23963,
      "observe": 23224,
      "factual": 10876,
      "correctness": 6205,
      "probing": 25979,
      "hidden": 13535,
      "edge": 8818,
      "serve": 29646,
      "dropin": 8741,
      "replacement": 28101,
      "taskrelated": 32223,
      "subword": 31501,
      "units": 34369,
      "morphologically": 22221,
      "rich": 28873,
      "proven": 26671,
      "particularly": 24345,
      "complexity": 5300,
      "makes": 19488,
      "apply": 2202,
      "studies": 31261,
      "considerable": 5703,
      "network": 22683,
      "ngrams": 22903,
      "general": 12154,
      "hungarian": 13948,
      "center": 4263,
      "transformergenerated": 33772,
      "works": 35809,
      "isolating": 15650,
      "causes": 4255,
      "explosion": 10624,
      "called": 3982,
      "subwordbased": 31503,
      "statistically": 31021,
      "derived": 7672,
      "bpe": 3848,
      "statistical": 31017,
      "tokenizers": 33212,
      "reducing": 27742,
      "memory": 19800,
      "requirements": 28235,
      "finally": 11187,
      "terms": 32736,
      "recognition": 27636,
      "words": 35655,
      "chess": 4532,
      "mastering": 19637,
      "play": 25016,
      "support": 31704,
      "generic": 12719,
      "strategic": 31097,
      "textarchived": 32968,
      "games": 12079,
      "addition": 1234,
      "skills": 30308,
      "abstract": 767,
      "meaningful": 19724,
      "moves": 22237,
      "gameplay": 12078,
      "28": 286,
      "million": 20162,
      "portable": 25149,
      "game": 12076,
      "notation": 23033,
      "steps": 31057,
      "optimizes": 23654,
      "weights": 35501,
      "generates": 12399,
      "displays": 8318,
      "formations": 11757,
      "identifiable": 13993,
      "classic": 4588,
      "exchange": 10181,
      "live": 18055,
      "interface": 15415,
      "correctly": 6202,
      "filters": 11174,
      "illegal": 14043,
      "anticipate": 2092,
      "build": 3907,
      "promise": 26273,
      "features": 11029,
      "underlying": 34152,
      "rule": 29006,
      "expressive": 10644,
      "player": 25022,
      "annotations": 2029,
      "navigating": 22594,
      "synthetic": 31845,
      "agents": 1566,
      "modern": 22155,
      "contain": 5825,
      "tremendous": 33881,
      "belief": 3336,
      "consistently": 5743,
      "accurately": 930,
      "reflect": 27788,
      "beliefs": 3337,
      "beings": 3334,
      "produced": 26156,
      "sociological": 30443,
      "tool": 33256,
      "ways": 35449,
      "distinct": 8362,
      "interviews": 15478,
      "surveys": 31779,
      "historical": 13725,
      "launch": 17417,
      "clusters": 4709,
      "strings": 31156,
      "orientation": 23695,
      "trajectories": 33658,
      "contained": 5831,
      "known": 15933,
      "ground": 13284,
      "truth": 33936,
      "board": 3801,
      "legality": 17699,
      "patterns": 24408,
      "piece": 24948,
      "creates": 6363,
      "accurate": 919,
      "latent": 17405,
      "representation": 28137,
      "plot": 25053,
      "legal": 17694,
      "compression": 5411,
      "survey": 31772,
      "fields": 11154,
      "ir": 15640,
      "recurrent": 27677,
      "networks": 22707,
      "rnns": 28912,
      "gated": 12124,
      "shortterm": 29828,
      "120": 92,
      "bidirectional": 3690,
      "encoder": 9345,
      "24": 262,
      "94": 599,
      "multitask": 22444,
      "73": 511,
      "134": 114,
      "95": 602,
      "98": 609,
      "gshard": 13324,
      "63": 468,
      "humongous": 13939,
      "applications": 2139,
      "demand": 7409,
      "small": 30335,
      "response": 28471,
      "times": 33156,
      "pruning": 26808,
      "quantization": 27003,
      "sharing": 29790,
      "tensor": 32721,
      "decomposition": 7289,
      "deployment": 7645,
      "industry": 14744,
      "critical": 6380,
      "need": 22616,
      "efficient": 9019,
      "published": 26870,
      "area": 2476,
      "plethora": 25035,
      "community": 5056,
      "story": 31086,
      "predictors": 25457,
      "page": 23964,
      "wellknown": 35525,
      "twofold": 34038,
      "firstly": 11564,
      "discriminate": 8282,
      "machinegenerated": 19368,
      "emerge": 9154,
      "detect": 7849,
      "content": 5848,
      "fast": 10990,
      "bootstrapping": 3831,
      "indicators": 14707,
      "secondly": 29333,
      "curious": 6478,
      "understand": 34186,
      "prevalence": 25855,
      "nature": 22588,
      "pages": 23966,
      "wild": 35607,
      "extensive": 10674,
      "qualitative": 26929,
      "quantitative": 26990,
      "500": 424,
      "web": 35471,
      "articles": 2523,
      "conducted": 5628,
      "topic": 33281,
      "comparative": 5095,
      "grading": 13199,
      "process": 26048,
      "student": 31252,
      "approaches": 2367,
      "desired": 7761,
      "implemented": 14167,
      "mapping": 19578,
      "facet": 10831,
      "conventional": 6071,
      "embeddings": 9137,
      "extracted": 10748,
      "elmo": 9112,
      "assess": 2586,
      "efficiency": 8995,
      "cosine": 6239,
      "similarity": 30126,
      "correlation": 6218,
      "measurements": 19740,
      "outperformed": 23791,
      "briefly": 3876,
      "conclude": 5550,
      "poor": 25101,
      "discriminator": 8288,
      "guided": 13352,
      "sequence": 29594,
      "imitate": 14106,
      "distribution": 8389,
      "realistic": 27315,
      "control": 6047,
      "regions": 27821,
      "especially": 9730,
      "problematic": 26020,
      "usually": 34944,
      "toxicity": 33315,
      "hate": 13471,
      "guide": 13342,
      "safer": 29039,
      "controllable": 6057,
      "guides": 13357,
      "step": 31036,
      "computing": 5511,
      "probabilities": 25968,
      "bayes": 3294,
      "normalizing": 23020,
      "distributions": 8402,
      "conditioned": 5575,
      "undesired": 34293,
      "stronger": 31196,
      "controllability": 6056,
      "achieving": 1079,
      "speeds": 30798,
      "faster": 10995,
      "just": 15733,
      "keyword": 15796,
      "unlocking": 34412,
      "lastly": 17398,
      "15b": 151,
      "toxic": 33307,
      "sacrificing": 29024,
      "far": 10986,
      "maintaining": 19418,
      "speed": 30794,
      "risks": 28894,
      "advanced": 1420,
      "expand": 10335,
      "potential": 25234,
      "abuse": 781,
      "assessing": 2608,
      "experimenting": 10414,
      "prompts": 26403,
      "representative": 28180,
      "narrative": 22491,
      "structures": 31230,
      "interaction": 15383,
      "ideologies": 14027,
      "predecessor": 25398,
      "gpt3s": 13048,
      "strength": 31138,
      "emulates": 9276,
      "interactive": 15393,
      "informational": 14927,
      "influential": 14846,
      "utilized": 34964,
      "individuals": 14722,
      "behaviors": 3325,
      "measures": 19741,
      "possibility": 25206,
      "technology": 32679,
      "represents": 28196,
      "absence": 758,
      "safeguards": 29037,
      "successful": 31531,
      "little": 18047,
      "experimentation": 10412,
      "likely": 17901,
      "stakeholders": 30862,
      "begin": 3305,
      "investing": 15617,
      "soon": 30522,
      "norms": 23023,
      "public": 26829,
      "educational": 8838,
      "initiatives": 14977,
      "influx": 14848,
      "disinformation": 8308,
      "propaganda": 26463,
      "mitigation": 20266,
      "government": 12839,
      "society": 30440,
      "current": 6479,
      "limitations": 17914,
      "classify": 4630,
      "tradeoff": 33334,
      "including": 14454,
      "extension": 10671,
      "batch": 3286,
      "length": 17703,
      "attention": 2711,
      "recurrence": 27676,
      "identify": 14002,
      "suffer": 31549,
      "example": 10108,
      "struggles": 31249,
      "loosely": 19224,
      "performing": 24833,
      "gpt23": 12971,
      "sim": 30094,
      "scaling": 29158,
      "efficiently": 9067,
      "argue": 2489,
      "reduce": 27698,
      "entire": 9622,
      "sample": 29063,
      "speculate": 30775,
      "modify": 22182,
      "perform": 24468,
      "causal": 4235,
      "retriever": 28777,
      "jointly": 15715,
      "extreme": 10790,
      "multilabel": 22290,
      "focus": 11644,
      "tail": 32013,
      "label": 15951,
      "insufficient": 15311,
      "mitigate": 20247,
      "takes": 32030,
      "advantage": 1497,
      "perturbations": 24910,
      "substantial": 31458,
      "contributions": 6039,
      "combine": 4956,
      "broad": 3883,
      "realtoxicityprompts": 27332,
      "evaluating": 9891,
      "degeneration": 7373,
      "prone": 26452,
      "racist": 27147,
      "sexist": 29756,
      "hinders": 13720,
      "safe": 29028,
      "extent": 10721,
      "prompted": 26366,
      "algorithms": 1717,
      "preventing": 25859,
      "100k": 50,
      "naturally": 22584,
      "occurring": 23277,
      "sentencelevel": 29548,
      "widelyused": 35578,
      "degenerate": 7372,
      "seemingly": 29362,
      "innocuous": 14990,
      "empirically": 9245,
      "adaptive": 1220,
      "nontoxic": 23012,
      "steering": 31028,
      "away": 3046,
      "solutions": 30483,
      "bad": 3072,
      "pinpoint": 24956,
      "cause": 4252,
      "persistent": 24876,
      "radford": 27149,
      "offensive": 23282,
      "factually": 10894,
      "unreliable": 34427,
      "bed": 3304,
      "selection": 29388,
      "processes": 26090,
      "capacities": 4123,
      "abilities": 622,
      "highquality": 13682,
      "discourse": 8263,
      "papers": 24147,
      "analyzed": 1999,
      "aspects": 2572,
      "encoded": 9338,
      "date": 7197,
      "intersentential": 15473,
      "quantitatively": 26996,
      "evaluates": 9887,
      "examine": 10098,
      "rhetoric": 28871,
      "encode": 9336,
      "theory": 33056,
      "revealing": 28809,
      "richer": 28876,
      "intermediate": 15423,
      "layer": 17423,
      "apparently": 2110,
      "suggest": 31566,
      "drawing": 8726,
      "avenue": 3004,
      "quantifying": 26988,
      "converting": 6114,
      "point": 25062,
      "view": 35273,
      "messages": 19854,
      "spoken": 30813,
      "virtual": 35283,
      "assistants": 2637,
      "quite": 27137,
      "literal": 18038,
      "user": 34644,
      "tell": 32684,
      "extract": 10743,
      "message": 19853,
      "contact": 5824,
      "named": 22480,
      "designed": 7718,
      "allow": 1800,
      "voice": 35385,
      "convert": 6112,
      "deliver": 7399,
      "target": 32045,
      "developed": 7922,
      "rulebased": 29007,
      "integrates": 15327,
      "linear": 17984,
      "partofspeech": 24365,
      "tagging": 32011,
      "parsing": 24319,
      "transformation": 33694,
      "investigated": 15601,
      "lstms": 19329,
      "copynet": 6146,
      "explored": 10607,
      "metrics": 20129,
      "gauge": 12130,
      "naturalness": 22587,
      "faithfulness": 10930,
      "automatically": 2901,
      "chose": 4561,
      "plus": 25059,
      "meteor": 19865,
      "separately": 29584,
      "slight": 30320,
      "830": 571,
      "159": 150,
      "37": 358,
      "publicly": 26848,
      "released": 27918,
      "composed": 5320,
      "samples": 29071,
      "incremental": 14655,
      "age": 1555,
      "encoders": 9377,
      "nlu": 22969,
      "humans": 13916,
      "incrementally": 14656,
      "assume": 2658,
      "processed": 26089,
      "behave": 3309,
      "interfaces": 15416,
      "partial": 24322,
      "seen": 29363,
      "certain": 4271,
      "happen": 13417,
      "mode": 20329,
      "retaining": 28718,
      "impacted": 14143,
      "alleviated": 1791,
      "adapting": 1206,
      "regime": 27815,
      "truncated": 33930,
      "procedure": 26044,
      "right": 28879,
      "incorporating": 14574,
      "hypothetical": 13971,
      "contexts": 5935,
      "semisupervised": 29504,
      "style": 31412,
      "indonesian": 14729,
      "informal": 14850,
      "formal": 11743,
      "iterative": 15687,
      "daily": 6577,
      "deviations": 7988,
      "standard": 30867,
      "spelling": 30804,
      "order": 23668,
      "typically": 34076,
      "parallel": 24165,
      "counterpart": 6301,
      "augmenting": 2822,
      "artificial": 2528,
      "dealing": 7209,
      "extremely": 10795,
      "transformerbased": 33745,
      "alternatively": 1857,
      "finedtuned": 11266,
      "equally": 9674,
      "costs": 6266,
      "resource": 28408,
      "findings": 11229,
      "promising": 26280,
      "leveraging": 17776,
      "ptlm": 26826,
      "nlg": 22909,
      "token": 33184,
      "bertstyle": 3551,
      "ptlms": 26827,
      "span": 30589,
      "infilling": 14834,
      "crucial": 6436,
      "contrastive": 6020,
      "taskspecific": 32559,
      "unify": 34346,
      "mutually": 22473,
      "reinforce": 27835,
      "calm": 3985,
      "relying": 27975,
      "graphs": 13238,
      "yielding": 35918,
      "relatively": 27884,
      "consistent": 5735,
      "margin": 19583,
      "suggests": 31590,
      "know": 15806,
      "calibration": 3976,
      "regarding": 27808,
      "facts": 10874,
      "perfect": 24466,
      "appropriate": 2405,
      "ask": 2551,
      "confidence": 5648,
      "property": 26479,
      "probabilistic": 25966,
      "predicted": 25413,
      "actually": 1156,
      "correlated": 6213,
      "qa": 26906,
      "calibrated": 3973,
      "finding": 11224,
      "calibrate": 3972,
      "correlate": 6211,
      "probability": 25970,
      "modification": 22178,
      "adjustment": 1393,
      "outputs": 23886,
      "inputs": 15040,
      "diverse": 8410,
      "strengths": 31141,
      "shedding": 29797,
      "light": 17826,
      "calibrating": 3974,
      "good": 12816,
      "successfully": 31535,
      "lag": 16013,
      "overcome": 23919,
      "dutch": 8755,
      "retraining": 28726,
      "lexical": 17798,
      "tuning": 33965,
      "aligned": 1736,
      "scale": 29128,
      "transforming": 33803,
      "medium": 19775,
      "embedding": 9127,
      "minimises": 20193,
      "prevents": 25863,
      "losing": 19239,
      "assessed": 2605,
      "par": 24149,
      "notoriously": 23047,
      "behavior": 3311,
      "recast": 27473,
      "controlling": 6067,
      "application": 2125,
      "apis": 2106,
      "programs": 26204,
      "altering": 1845,
      "hyperparameters": 13960,
      "paradigm": 24154,
      "specialized": 30664,
      "manipulating": 19542,
      "activations": 1142,
      "produce": 26137,
      "changes": 4409,
      "allowing": 1804,
      "contribute": 6027,
      "construction": 5810,
      "algorithm": 1699,
      "loss": 19240,
      "function": 11960,
      "efficacy": 8985,
      "noun": 23049,
      "aversion": 3032,
      "speech": 30779,
      "filtering": 11172,
      "largely": 17306,
      "controlled": 6062,
      "fluency": 11630,
      "deterministic": 7903,
      "settings": 29728,
      "uncertainty": 34115,
      "surprisal": 31761,
      "exploiting": 10558,
      "humor": 13940,
      "studied": 31259,
      "datadriven": 6925,
      "actual": 1155,
      "mechanism": 19748,
      "break": 3858,
      "components": 5310,
      "setup": 29748,
      "special": 30654,
      "relationship": 27872,
      "developing": 7939,
      "audience": 2773,
      "expectations": 10350,
      "increasingly": 14632,
      "feed": 11054,
      "calculate": 3961,
      "values": 35026,
      "conducting": 5644,
      "semeval": 29490,
      "2021": 228,
      "event": 10045,
      "schema": 29230,
      "temporal": 32693,
      "relationships": 27874,
      "addresses": 1361,
      "ordering": 23679,
      "sorting": 30529,
      "occurred": 23274,
      "predicting": 25416,
      "fit": 11571,
      "bartbased": 3111,
      "temporality": 32701,
      "cooccurrence": 6127,
      "meaning": 19718,
      "flexibly": 11618,
      "denoising": 7600,
      "shuffle": 29939,
      "delete": 7394,
      "recover": 27666,
      "inferences": 14825,
      "incomplete": 14537,
      "outperforming": 23795,
      "pairwise": 23986,
      "pointer": 25067,
      "temporally": 32702,
      "pile": 24952,
      "crossdomain": 6407,
      "825": 567,
      "targeted": 32062,
      "22": 250,
      "subsets": 31455,
      "newly": 22868,
      "derive": 7670,
      "academic": 784,
      "professional": 26174,
      "sources": 30574,
      "untuned": 34464,
      "conversely": 6111,
      "raw": 27281,
      "cc": 4260,
      "evaluations": 10026,
      "indepth": 14672,
      "exploratory": 10567,
      "potentially": 25309,
      "concerning": 5539,
      "lottery": 19262,
      "heavily": 13492,
      "enormous": 9585,
      "focusing": 11670,
      "requiring": 28267,
      "expense": 10355,
      "higher": 13593,
      "demands": 7413,
      "computer": 5502,
      "vision": 35291,
      "computationallyefficient": 5489,
      "applicable": 2124,
      "inside": 15062,
      "winning": 35616,
      "early": 8773,
      "stage": 30857,
      "comprehensive": 5350,
      "squad": 30841,
      "prefixtuning": 25485,
      "optimizing": 23655,
      "continuous": 5995,
      "facto": 10861,
      "modifies": 22181,
      "necessitates": 22612,
      "storing": 31085,
      "copy": 6144,
      "lightweight": 17836,
      "keeps": 15747,
      "frozen": 11935,
      "vector": 35193,
      "prefix": 25479,
      "draws": 8730,
      "inspiration": 15089,
      "prompting": 26371,
      "subsequent": 31445,
      "attend": 2710,
      "tabletotext": 31989,
      "summarization": 31608,
      "01": 2,
      "obtains": 23262,
      "lowdata": 19278,
      "switch": 31795,
      "trillion": 33902,
      "sparsity": 30628,
      "reuse": 28785,
      "mixture": 20279,
      "experts": 10518,
      "moe": 22195,
      "selects": 29401,
      "outrageous": 23899,
      "constant": 5769,
      "cost": 6242,
      "despite": 7767,
      "notable": 23024,
      "successes": 31530,
      "widespread": 35589,
      "adoption": 1407,
      "hindered": 13715,
      "communication": 5048,
      "instability": 15106,
      "simplify": 30180,
      "routing": 28994,
      "design": 7695,
      "intuitive": 15558,
      "reduced": 27728,
      "help": 13501,
      "instabilities": 15105,
      "sparse": 30609,
      "lower": 19282,
      "precision": 25391,
      "bfloat16": 3639,
      "formats": 11758,
      "t5base": 31969,
      "t5large": 31975,
      "7x": 550,
      "increases": 14613,
      "multilingual": 22294,
      "mt5base": 22258,
      "101": 53,
      "advance": 1416,
      "colossal": 4945,
      "clean": 4643,
      "crawled": 6345,
      "achieve": 941,
      "4x": 416,
      "speedup": 30800,
      "t5xxl": 31981,
      "impact": 14119,
      "phrase": 24936,
      "suggestions": 31586,
      "email": 9121,
      "composition": 5325,
      "behaviour": 3331,
      "native": 22498,
      "nonnative": 23003,
      "writers": 35847,
      "multiword": 22466,
      "suggestion": 31585,
      "choices": 4556,
      "compares": 5185,
      "vs": 35400,
      "ideation": 13990,
      "emerging": 9190,
      "literature": 18040,
      "editor": 8831,
      "prototype": 26664,
      "refined": 27780,
      "people": 24450,
      "emails": 9122,
      "conditions": 5578,
      "benefits": 3489,
      "phrases": 24938,
      "speakers": 30649,
      "insights": 15068,
      "implications": 14172,
      "supporting": 31715,
      "replacing": 28105,
      "active": 1144,
      "platform": 25008,
      "managed": 19528,
      "service": 29657,
      "utilizes": 34966,
      "business": 3944,
      "quickly": 27135,
      "easily": 8790,
      "deploy": 7629,
      "ready": 27308,
      "hosted": 13761,
      "environment": 9665,
      "involvement": 15626,
      "scientists": 29262,
      "leverages": 17765,
      "implementation": 14160,
      "workflow": 35801,
      "relies": 27965,
      "providing": 26769,
      "labeling": 15961,
      "experience": 10371,
      "reallife": 27324,
      "insurance": 15312,
      "ideal": 13986,
      "scripts": 29296,
      "vanilla": 35028,
      "adjustments": 1395,
      "targeting": 32066,
      "encountered": 9395,
      "practice": 25377,
      "list": 18033,
      "plan": 24993,
      "solve": 30487,
      "presented": 25565,
      "script": 29294,
      "planned": 24996,
      "examplebased": 10115,
      "prompt": 26307,
      "onthefly": 23373,
      "domains": 8610,
      "incredible": 14651,
      "outofdistribution": 23746,
      "underexplored": 34140,
      "unknown": 34378,
      "unlabeled": 34379,
      "respect": 28450,
      "unrestricted": 34429,
      "consisting": 5759,
      "characterize": 4429,
      "intuitively": 15560,
      "signature": 29945,
      "maps": 19580,
      "multisource": 22439,
      "scenarios": 29199,
      "semantically": 29482,
      "puzzles": 26898,
      "dominant": 8659,
      "advancing": 1491,
      "seek": 29357,
      "highly": 13656,
      "read": 27293,
      "fluent": 11632,
      "adversarially": 1520,
      "parts": 24369,
      "definition": 7368,
      "characterlevel": 4434,
      "manipulations": 19545,
      "expert": 10506,
      "creative": 6370,
      "intelligence": 15350,
      "combining": 4961,
      "main": 19386,
      "humanlike": 13906,
      "contribution": 6036,
      "curriculum": 6547,
      "split": 30810,
      "perturbing": 24913,
      "exhibits": 10242,
      "partially": 24323,
      "considerably": 5711,
      "bestperforming": 3582,
      "fails": 10913,
      "generalize": 12231,
      "remain": 27982,
      "unsolved": 34447,
      "innovation": 14991,
      "predict": 25404,
      "fall": 10947,
      "predetermined": 25401,
      "categories": 4221,
      "classifications": 4621,
      "utilize": 34958,
      "reported": 28126,
      "rarely": 27259,
      "occur": 23273,
      "created": 6358,
      "lines": 18000,
      "practitioners": 25381,
      "flipping": 11620,
      "images": 14082,
      "increase": 14591,
      "volume": 35386,
      "image": 14054,
      "convolutional": 6122,
      "purpose": 26882,
      "creating": 6364,
      "utilizing": 34968,
      "yelp": 35901,
      "restaurant": 28527,
      "reviews": 28832,
      "combined": 4958,
      "genuine": 12723,
      "extractive": 10777,
      "abstractive": 776,
      "explanations": 10540,
      "factchecking": 10858,
      "news": 22874,
      "claims": 4579,
      "assisting": 2639,
      "experiment": 10376,
      "biased": 3668,
      "graphbased": 13232,
      "misinformation": 20212,
      "political": 25093,
      "health": 13483,
      "plain": 24989,
      "outstanding": 23902,
      "mental": 19838,
      "states": 31010,
      "emotional": 9203,
      "variations": 35048,
      "extensions": 10673,
      "optimize": 23640,
      "leaderboard": 17471,
      "viable": 35241,
      "rewarding": 28862,
      "formality": 11747,
      "scarcity": 29188,
      "scarce": 29187,
      "preserving": 25608,
      "boosts": 3827,
      "preservation": 25601,
      "rewards": 28863,
      "core": 6149,
      "appealing": 2111,
      "rely": 27968,
      "popular": 25111,
      "solution": 30467,
      "candidate": 3990,
      "affected": 1540,
      "irrelevant": 15643,
      "factors": 10866,
      "frequencies": 11922,
      "distracting": 8380,
      "mislead": 20216,
      "choose": 4558,
      "wrong": 35865,
      "oversensitive": 23952,
      "correct": 6190,
      "considering": 5722,
      "devise": 7992,
      "sound": 30542,
      "formalism": 11746,
      "verify": 35216,
      "robustness": 28942,
      "attacked": 2689,
      "synonym": 31809,
      "drops": 8746,
      "indicating": 14701,
      "timedial": 33151,
      "dialog": 8003,
      "conversations": 6105,
      "turn": 34029,
      "massive": 19622,
      "dialogs": 8008,
      "introducing": 15549,
      "formulate": 11770,
      "multiplechoice": 22429,
      "cloze": 4705,
      "carefully": 4170,
      "23": 256,
      "absolute": 760,
      "shallow": 29773,
      "motivating": 22231,
      "robust": 28929,
      "contextual": 5948,
      "bottleneck": 3838,
      "intensive": 15370,
      "involved": 15625,
      "decoding": 7268,
      "accelerate": 790,
      "optimization": 23621,
      "cache": 3957,
      "detecting": 7854,
      "repeated": 28090,
      "asynchronous": 2677,
      "pipeline": 24962,
      "io": 15636,
      "optimizations": 23639,
      "gain": 12055,
      "easy": 8800,
      "oneline": 23342,
      "change": 4406,
      "costeffective": 6258,
      "grown": 13319,
      "leaps": 17504,
      "bounds": 3846,
      "limit": 17907,
      "utilization": 34952,
      "suite": 31598,
      "implement": 14156,
      "toolkit": 33265,
      "encoderdecoder": 9362,
      "bilingual": 3706,
      "billion": 3712,
      "198": 195,
      "mt5": 22254,
      "excellent": 10160,
      "having": 13472,
      "tens": 32715,
      "billions": 3726,
      "gpu": 13167,
      "ernie": 9705,
      "175": 171,
      "taskagnostic": 32212,
      "zeroshotfewshot": 35999,
      "kind": 15800,
      "unified": 34323,
      "fuses": 12019,
      "autoencoding": 2846,
      "graph": 13217,
      "54": 441,
      "chinese": 4540,
      "place": 24983,
      "superglue": 31641,
      "july": 15730,
      "08": 16,
      "spanish": 30593,
      "family": 10973,
      "includes": 14449,
      "robertabase": 28925,
      "robertalarge": 28926,
      "gpt2large": 12973,
      "arguably": 2488,
      "proficient": 26185,
      "deduplicated": 7312,
      "135": 116,
      "archive": 2475,
      "national": 22496,
      "ex": 10083,
      "novo": 23130,
      "picard": 24944,
      "fictional": 11130,
      "introduced": 15539,
      "star": 30893,
      "dictionary": 8029,
      "construct": 5793,
      "456": 395,
      "76": 518,
      "translating": 33816,
      "rethinking": 28723,
      "supplementary": 31702,
      "technique": 32614,
      "finetunes": 11364,
      "involving": 15634,
      "discover": 8266,
      "orthogonal": 23732,
      "discrimination": 8284,
      "synthesized": 31840,
      "role": 28952,
      "headlines": 13479,
      "causalities": 4248,
      "implicit": 14175,
      "relations": 27869,
      "5000": 425,
      "headline": 13478,
      "russian": 29019,
      "crowdsourcing": 6434,
      "vary": 35164,
      "totally": 33301,
      "unrelated": 34425,
      "belonging": 3346,
      "ones": 23344,
      "validity": 35008,
      "xlmroberta": 35877,
      "causality": 4249,
      "detection": 7861,
      "intent": 15371,
      "noisy": 22986,
      "enhance": 9500,
      "offtopic": 23333,
      "mining": 20202,
      "oneshot": 23350,
      "expanded": 10338,
      "teaching": 32590,
      "demonstration": 7594,
      "gptneo": 13155,
      "appropriately": 2409,
      "stepbystep": 31053,
      "demonstrations": 7596,
      "teach": 32578,
      "execute": 10190,
      "mathematical": 19677,
      "previously": 25896,
      "proved": 26668,
      "operations": 23568,
      "mathematics": 19686,
      "200": 216,
      "division": 8492,
      "reporting": 28127,
      "smallest": 30404,
      "80": 552,
      "constructing": 5808,
      "coax": 4717,
      "kinds": 15801,
      "multistep": 22441,
      "tremendously": 33885,
      "numerical": 23177,
      "required": 28228,
      "reasons": 27467,
      "learn": 17505,
      "predecessors": 25399,
      "consider": 5701,
      "magnitude": 19381,
      "minimum": 20201,
      "reasonably": 27363,
      "interpolation": 15452,
      "extrapolation": 10789,
      "autocompletion": 2841,
      "syntactically": 31829,
      "effort": 9077,
      "adapt": 1162,
      "medical": 19766,
      "encourages": 9400,
      "complete": 5248,
      "enriched": 9592,
      "eventually": 10049,
      "preliminary": 25487,
      "understood": 34280,
      "ambiguous": 1874,
      "ambiguities": 1872,
      "arise": 2495,
      "beginning": 3306,
      "compatible": 5203,
      "exhibit": 10210,
      "modulated": 22190,
      "probe": 25976,
      "stochastic": 31068,
      "completions": 5264,
      "estimate": 9780,
      "assigns": 2629,
      "interpretation": 15462,
      "parses": 24318,
      "unlike": 34392,
      "hypothesized": 13970,
      "researcher": 28369,
      "lstm": 19326,
      "materials": 19664,
      "simultaneously": 30194,
      "varies": 35052,
      "constructions": 5812,
      "occasional": 23267,
      "areas": 2484,
      "truthfulqa": 33943,
      "measuring": 19742,
      "mimic": 20169,
      "falsehoods": 10962,
      "truthful": 33937,
      "38": 359,
      "law": 17418,
      "finance": 11208,
      "politics": 25096,
      "crafted": 6340,
      "false": 10957,
      "misconception": 20210,
      "imitating": 14107,
      "t5based": 31972,
      "58": 452,
      "misconceptions": 20211,
      "deceive": 7222,
      "generally": 12239,
      "contrasts": 6026,
      "expected": 10351,
      "truthfulness": 33939,
      "imitation": 14109,
      "ner": 22676,
      "numerous": 23181,
      "sota": 30530,
      "relevant": 27935,
      "necessarily": 22605,
      "reports": 28128,
      "traded": 33332,
      "companies": 5068,
      "combines": 4960,
      "inferencing": 14828,
      "220m": 251,
      "rougel": 28982,
      "chatgpt": 4455,
      "highlighting": 13642,
      "difficulty": 8176,
      "49": 409,
      "surprise": 31763,
      "financespecific": 11215,
      "llm": 18259,
      "15": 140,
      "artifacts": 2527,
      "encourage": 9397,
      "direction": 8221,
      "sophisticated": 30524,
      "financial": 11216,
      "relating": 27862,
      "exposure": 10635,
      "focuses": 11665,
      "mentioned": 19843,
      "remedy": 28069,
      "verified": 35212,
      "holtzman": 13746,
      "2020": 227,
      "motivated": 22227,
      "relate": 27852,
      "qualitatively": 26936,
      "mistakes": 20226,
      "occurs": 23278,
      "significance": 29947,
      "inspecting": 15086,
      "partly": 24363,
      "caused": 4254,
      "explaining": 10533,
      "amplify": 1896,
      "sum": 31604,
      "concrete": 5558,
      "foundation": 11790,
      "investigation": 15612,
      "finegrained": 11267,
      "annotated": 2015,
      "fixed": 11577,
      "extending": 10662,
      "classes": 4587,
      "accommodate": 837,
      "coarsetofine": 4716,
      "grained": 13202,
      "asking": 2558,
      "opt": 23586,
      "surface": 31723,
      "names": 22490,
      "guidance": 13341,
      "formulation": 11776,
      "generators": 12718,
      "regularization": 27827,
      "constraints": 5787,
      "giving": 12785,
      "prior": 25934,
      "bootstraps": 3832,
      "refinement": 27781,
      "case": 4193,
      "superior": 31643,
      "kronecker": 15942,
      "attracted": 2751,
      "lot": 19259,
      "attributed": 2759,
      "huge": 13778,
      "100m": 52,
      "prohibitive": 26235,
      "deploying": 7638,
      "devices": 7990,
      "mitigated": 20260,
      "compressing": 5409,
      "compress": 5401,
      "mappings": 19579,
      "initialized": 14969,
      "decomposed": 7285,
      "undergone": 34147,
      "portion": 25150,
      "distilgpt2": 8332,
      "decoderbased": 7249,
      "wide": 35544,
      "increasing": 14616,
      "encoderbased": 9361,
      "tinybert": 33174,
      "distilbert": 8330,
      "distilroberta": 8360,
      "employ": 9253,
      "compressed": 5404,
      "truncation": 33931,
      "distillationbased": 8348,
      "cleaning": 4645,
      "emerged": 9155,
      "splits": 30811,
      "tuned": 33959,
      "t5xl": 31980,
      "ablation": 733,
      "minimization": 20194,
      "engineering": 9461,
      "efforts": 9083,
      "capacity": 4124,
      "comparatively": 5102,
      "sam": 29062,
      "convergence": 6083,
      "flatter": 11611,
      "minima": 20180,
      "overhead": 23935,
      "trivia": 33918,
      "tydiqa": 34051,
      "product": 26167,
      "ecommerce": 8808,
      "recognize": 27644,
      "amazon": 1868,
      "alexa": 1697,
      "items": 15680,
      "utterance": 34978,
      "perfectly": 24467,
      "ngram": 22900,
      "helps": 13522,
      "customers": 6559,
      "shopping": 29808,
      "central": 4266,
      "serving": 29663,
      "starting": 30899,
      "pain": 23968,
      "persist": 24875,
      "grow": 13309,
      "bigger": 3703,
      "175b": 175,
      "timeconsuming": 33147,
      "default": 7356,
      "sensible": 29511,
      "functionality": 11967,
      "deployed": 7632,
      "resourceconstrained": 28418,
      "environments": 9668,
      "parameterefficient": 24205,
      "weight": 35493,
      "updates": 34476,
      "final": 11175,
      "dubbed": 8751,
      "enforcing": 9451,
      "sparsityaware": 30636,
      "lowrank": 19299,
      "resourceefficient": 28423,
      "encouraging": 9401,
      "investigations": 15616,
      "backbones": 3057,
      "dozens": 8704,
      "instance": 15108,
      "saves": 29115,
      "25": 270,
      "flops": 11625,
      "05": 11,
      "trainable": 33379,
      "codes": 4848,
      "underpin": 34168,
      "introduction": 15554,
      "contributed": 6031,
      "advancements": 1459,
      "field": 11132,
      "come": 4965,
      "grows": 13320,
      "quadratically": 26928,
      "extends": 10668,
      "childrens": 4536,
      "book": 3809,
      "guiding": 13358,
      "enhancing": 9555,
      "tackling": 32005,
      "applicability": 2117,
      "specialised": 30658,
      "instances": 15111,
      "aim": 1636,
      "analyse": 1905,
      "seed": 29351,
      "gptgenerated": 13144,
      "consequently": 5695,
      "exploit": 10554,
      "hierarchical": 13541,
      "handful": 13404,
      "opens": 23477,
      "interesting": 15406,
      "avenues": 3005,
      "blockwise": 3780,
      "scalable": 29123,
      "enhancement": 9540,
      "residual": 28388,
      "scheme": 29232,
      "internal": 15436,
      "blocks": 3777,
      "sequentially": 29629,
      "scalability": 29122,
      "lets": 17726,
      "dynamically": 8766,
      "runtime": 29016,
      "depending": 7619,
      "modularize": 22188,
      "needs": 22651,
      "incurring": 14661,
      "minimal": 20181,
      "added": 1229,
      "degradation": 7374,
      "novelty": 23126,
      "copying": 6145,
      "generalizable": 12202,
      "abstractions": 775,
      "tease": 32601,
      "apart": 2100,
      "possibilities": 25204,
      "sequential": 29624,
      "transformerxl": 33801,
      "local": 19127,
      "individual": 14715,
      "modelgenerated": 20882,
      "humangenerated": 13899,
      "largerscale": 17340,
      "1000": 42,
      "gpt2s": 12976,
      "frequent": 11924,
      "selfcontradictory": 29415,
      "ideally": 13987,
      "challenges": 4332,
      "slow": 30331,
      "sparsifying": 30627,
      "searching": 29316,
      "mask": 19606,
      "discrete": 8279,
      "matrices": 19689,
      "insight": 15064,
      "superset": 31669,
      "products": 26172,
      "hardware": 13427,
      "block": 3776,
      "flat": 11610,
      "pattern": 24406,
      "sparsify": 30626,
      "mlp": 20303,
      "3x": 372,
      "favorable": 11010,
      "accuracyefficiency": 917,
      "tradeoffs": 33338,
      "imagenet": 14079,
      "wikitext103": 35606,
      "25x": 276,
      "dense": 7604,
      "drop": 8739,
      "shifting": 29806,
      "generalpurpose": 12245,
      "methodology": 19991,
      "meet": 19779,
      "hallucinations": 13386,
      "violations": 35281,
      "raises": 27166,
      "representing": 28194,
      "energybased": 9449,
      "approximating": 2417,
      "gradients": 13198,
      "unconditional": 34130,
      "proposing": 26631,
      "robustly": 28941,
      "meeting": 19782,
      "contrast": 6010,
      "databased": 6919,
      "curricula": 6546,
      "lots": 19261,
      "acquisition": 1126,
      "researchers": 28370,
      "sequencing": 29623,
      "taskbased": 32214,
      "sampling": 29090,
      "investigates": 15604,
      "random": 27174,
      "initialization": 14968,
      "monolingual": 22208,
      "exceedingly": 10150,
      "alleviate": 1788,
      "tokenization": 33209,
      "replaced": 28100,
      "static": 31012,
      "covering": 6322,
      "french": 11921,
      "german": 12730,
      "accessible": 832,
      "damaging": 6582,
      "remarkable": 28024,
      "driven": 8735,
      "heterogeneous": 13530,
      "transferring": 33690,
      "continuing": 5992,
      "overlapping": 23941,
      "represent": 28131,
      "tree": 33877,
      "node": 22981,
      "avoiding": 3039,
      "negative": 22658,
      "interference": 15418,
      "100": 32,
      "represented": 28190,
      "websites": 35489,
      "c4": 3956,
      "heldout": 13498,
      "averaging": 3031,
      "paths": 24400,
      "marginal": 19587,
      "latency": 17400,
      "adjusting": 1392,
      "desirable": 7757,
      "proposes": 26624,
      "adjusts": 1396,
      "adaptively": 1222,
      "phase": 24916,
      "detects": 7889,
      "elements": 9100,
      "eliminates": 9107,
      "acc": 789,
      "metric": 20118,
      "adjusted": 1391,
      "selections": 29397,
      "bertbase": 3538,
      "sentiment": 29565,
      "regression": 27822,
      "global": 12794,
      "mathematically": 19685,
      "experimentally": 10411,
      "48": 404,
      "075": 15,
      "passable": 24377,
      "suggested": 31580,
      "posits": 25200,
      "llms": 18397,
      "necessary": 22606,
      "assemble": 2583,
      "software": 30452,
      "requirement": 28234,
      "undergoing": 34146,
      "shift": 29804,
      "surpass": 31732,
      "specially": 30679,
      "customized": 6564,
      "trend": 33886,
      "reusing": 28787,
      "flexible": 11616,
      "codebert": 4836,
      "gaussian": 12131,
      "noise": 22984,
      "simulation": 30190,
      "schemes": 29234,
      "perspective": 24895,
      "comparing": 5187,
      "keyphrase": 15790,
      "dedicated": 7303,
      "paradigms": 24161,
      "simplicity": 30170,
      "commonly": 5023,
      "adapted": 1197,
      "exposing": 10634,
      "limits": 17979,
      "parity": 24312,
      "adversarial": 1513,
      "lack": 15973,
      "players": 25023,
      "rival": 28896,
      "extra": 10737,
      "engagement": 9454,
      "designer": 7749,
      "collected": 4919,
      "collect": 4911,
      "yesno": 35903,
      "11b": 84,
      "702": 500,
      "941": 601,
      "unifiedskg": 34342,
      "unifying": 34347,
      "multitasking": 22457,
      "grounding": 13291,
      "requests": 28208,
      "databases": 6920,
      "bases": 3277,
      "communities": 5055,
      "systematic": 31865,
      "limitation": 17909,
      "format": 11753,
      "aiming": 1655,
      "promote": 26300,
      "exclusive": 10186,
      "modifications": 22179,
      "facilitates": 10849,
      "t0": 31931,
      "codex": 4861,
      "encoding": 9381,
      "extensible": 10670,
      "opensourced": 23547,
      "fairness": 10926,
      "receiving": 27485,
      "effect": 8849,
      "reduction": 27761,
      "interpreted": 15465,
      "line": 17981,
      "describes": 7680,
      "discussion": 8299,
      "hints": 13724,
      "fairer": 10924,
      "books": 3810,
      "article": 2520,
      "explores": 10612,
      "technologies": 32676,
      "stated": 30912,
      "replace": 28099,
      "publications": 26846,
      "tools": 33267,
      "ideas": 13989,
      "stands": 30890,
      "precisely": 25390,
      "supported": 31714,
      "value": 35022,
      "discussed": 8296,
      "emphasizes": 9212,
      "artistic": 2545,
      "issue": 15652,
      "comes": 4969,
      "aigenerated": 1632,
      "introduces": 15541,
      "projects": 26255,
      "artist": 2544,
      "andor": 2009,
      "deeplearningbased": 7350,
      "selecting": 29384,
      "suitable": 31596,
      "essential": 9756,
      "ml": 20289,
      "strongly": 31206,
      "primarily": 25907,
      "cv": 6572,
      "error": 9707,
      "iii": 14042,
      "compute": 5492,
      "huggingface": 13786,
      "hundreds": 13942,
      "systematically": 31876,
      "51": 430,
      "families": 10964,
      "niche": 22904,
      "status": 31024,
      "heavytail": 13496,
      "exhibiting": 10240,
      "correlations": 6221,
      "formulations": 11778,
      "pl": 24982,
      "spectral": 30769,
      "exponential": 10627,
      "exp": 10334,
      "maximizing": 19704,
      "adam": 1160,
      "1bit": 200,
      "gradient": 13185,
      "drastic": 8720,
      "distributed": 8384,
      "sgd": 29770,
      "nonlinearity": 23002,
      "individually": 14721,
      "optimizer": 23649,
      "estimates": 9783,
      "adaptivity": 1223,
      "wallclock": 35412,
      "guarantee": 13329,
      "smooth": 30410,
      "nonconvex": 22989,
      "bertlarge": 3543,
      "128": 101,
      "gpus": 13178,
      "87": 580,
      "rounds": 28987,
      "2times": 302,
      "throughput": 33098,
      "enjoying": 9577,
      "variational": 35045,
      "seminal": 29496,
      "originally": 23726,
      "generalized": 12235,
      "glm": 12790,
      "procedures": 26046,
      "converge": 6082,
      "offer": 23286,
      "freezing": 11920,
      "gnn": 12804,
      "cnn": 4711,
      "descent": 7676,
      "mixtureofexperts": 20284,
      "keeping": 15745,
      "unchanged": 34121,
      "resulting": 28552,
      "load": 19126,
      "allocates": 1798,
      "topk": 33293,
      "regardless": 27812,
      "importance": 14185,
      "employing": 9257,
      "variable": 35032,
      "speedups": 30803,
      "top1": 33277,
      "top2": 33278,
      "gating": 12129,
      "2x": 303,
      "selected": 29381,
      "activation": 1138,
      "super": 31636,
      "swift": 31791,
      "capable": 4105,
      "great": 13247,
      "incur": 14657,
      "separate": 29582,
      "fixedsize": 11581,
      "lose": 19237,
      "heavy": 13495,
      "decision": 7232,
      "routes": 28992,
      "energy": 9446,
      "agnostic": 1587,
      "blackbox": 3753,
      "architectural": 2433,
      "modules": 22194,
      "encoderonly": 9374,
      "wmt": 35629,
      "computations": 5491,
      "33times": 334,
      "29times": 293,
      "demo": 7419,
      "promptbased": 26358,
      "prompttuning": 26451,
      "hypernetworks": 13957,
      "learnable": 17516,
      "hypernetwork": 13955,
      "memories": 19791,
      "performances": 24819,
      "extended": 10659,
      "adopt": 1400,
      "matrix": 19692,
      "operator": 23575,
      "physics": 24942,
      "reconstruct": 27658,
      "specificity": 30764,
      "tensors": 32725,
      "illustrated": 14048,
      "memorize": 19794,
      "reproduce": 28198,
      "paraphrased": 24302,
      "contextually": 5968,
      "plagiarism": 24987,
      "verbatim": 35205,
      "paraphrase": 24300,
      "comparison": 5193,
      "domainspecific": 8645,
      "extensively": 10716,
      "exist": 10257,
      "memorization": 19793,
      "degrees": 7391,
      "majority": 19445,
      "scraped": 29287,
      "informing": 14936,
      "owners": 23959,
      "ethical": 9802,
      "exacerbate": 10084,
      "raising": 27171,
      "indiscriminately": 14710,
      "pursuing": 26887,
      "personal": 24880,
      "sensitive": 29513,
      "practicality": 25375,
      "urge": 34492,
      "discussions": 8302,
      "phenomena": 24921,
      "relied": 27963,
      "elaborate": 9093,
      "rnn": 28911,
      "dependency": 7617,
      "dependent": 7618,
      "nexttoken": 22893,
      "mixing": 20272,
      "modular": 22185,
      "responses": 28484,
      "modularity": 22187,
      "zhou": 36002,
      "internet": 15447,
      "applies": 2201,
      "blenderbot": 3761,
      "chen": 4528,
      "knowledgegrounded": 15927,
      "consistency": 5730,
      "engagingness": 9457,
      "topical": 33287,
      "brown": 3896,
      "factuality": 10890,
      "topicality": 33288,
      "vastly": 35191,
      "inducing": 14734,
      "anomalies": 2035,
      "deliberate": 7396,
      "dl": 8493,
      "delivered": 7400,
      "discriminating": 8283,
      "cognitively": 4882,
      "healthy": 13488,
      "alzheimers": 1862,
      "disease": 8303,
      "fitting": 11574,
      "artificially": 2541,
      "degraded": 7381,
      "ratio": 27274,
      "impaired": 14149,
      "theft": 33044,
      "description": 7682,
      "established": 9772,
      "alternatives": 1858,
      "generalizes": 12236,
      "spontaneous": 30818,
      "demonstrating": 7579,
      "induction": 14735,
      "inner": 14987,
      "workings": 35805,
      "dementia": 7418,
      "continually": 5981,
      "unfamiliar": 34307,
      "innovative": 14994,
      "employs": 9262,
      "initially": 14972,
      "subsequently": 31448,
      "evaluated": 9871,
      "29": 292,
      "outpaced": 23760,
      "enriches": 9593,
      "marks": 19604,
      "feedforward": 11075,
      "predictions": 25440,
      "promoting": 26303,
      "unveiling": 34468,
      "reverseengineering": 28825,
      "operation": 23565,
      "ffn": 11127,
      "additive": 1305,
      "update": 34473,
      "vectors": 35197,
      "humaninterpretable": 13900,
      "exit": 10331,
      "saving": 29116,
      "networkbased": 22705,
      "complicated": 5306,
      "distribute": 8383,
      "supercomputer": 31639,
      "tpus": 33322,
      "prevent": 25858,
      "bottlenecks": 3841,
      "reproducible": 28203,
      "libraries": 17811,
      "ease": 8787,
      "simplifies": 30179,
      "api": 2101,
      "creation": 6369,
      "pipelines": 24972,
      "opensource": 23482,
      "terabytes": 32728,
      "gptlike": 13153,
      "decoderonly": 7251,
      "excel": 10153,
      "fourier": 11811,
      "transform": 33693,
      "unfavorable": 34308,
      "tractable": 33330,
      "approximate": 2411,
      "hardwareefficient": 13435,
      "parameterized": 24218,
      "transforms": 33805,
      "surprisingly": 31766,
      "analytical": 1982,
      "optimal": 23611,
      "properties": 26470,
      "unlock": 34409,
      "vit": 35368,
      "reconstruction": 27660,
      "reverse": 28822,
      "sparsification": 30625,
      "openwebtext": 23559,
      "brings": 3879,
      "optimized": 23644,
      "nvidia": 23190,
      "mlperf": 20305,
      "record": 27662,
      "proofofconcept": 26461,
      "approximation": 2418,
      "17x": 182,
      "infused": 14940,
      "recalling": 27470,
      "tend": 32703,
      "counterfactual": 6294,
      "hallucinatory": 13399,
      "knowledgeintensive": 15929,
      "remedies": 28068,
      "modifying": 22183,
      "normally": 23021,
      "costly": 6263,
      "maintain": 19413,
      "interacting": 15382,
      "continuously": 6001,
      "constraint": 5784,
      "reinforcement": 27837,
      "armed": 2507,
      "seven": 29751,
      "confirms": 5663,
      "alleviates": 1792,
      "stable": 30848,
      "longer": 19194,
      "restoration": 28529,
      "abstraction": 774,
      "working": 35803,
      "simulates": 30188,
      "omitted": 23336,
      "identifies": 14001,
      "soft": 30447,
      "annotation": 2024,
      "generator": 12716,
      "gptneox20b": 13162,
      "freely": 11914,
      "openly": 23472,
      "permissive": 24862,
      "license": 17814,
      "submission": 31430,
      "languageunderstanding": 16924,
      "knowledgebased": 15926,
      "reasoner": 27365,
      "fiveshot": 11575,
      "sized": 30291,
      "learners": 17525,
      "zero": 35935,
      "13": 106,
      "60": 458,
      "wikipedia": 35602,
      "deepspeed": 7354,
      "megatron": 19786,
      "frameworks": 11907,
      "xglm": 35869,
      "facebook": 10825,
      "countries": 6305,
      "motivation": 22233,
      "thoroughly": 33074,
      "preparation": 25497,
      "versions": 35234,
      "covered": 6320,
      "xl": 35873,
      "supernaturalinstructions": 31665,
      "declarative": 7242,
      "expertwritten": 10522,
      "covers": 6330,
      "rigorous": 28883,
      "benchmarking": 3421,
      "crosstask": 6426,
      "follow": 11675,
      "subset": 31449,
      "remaining": 27990,
      "tkinstruct": 33180,
      "incontext": 14542,
      "definitions": 7369,
      "kshot": 15943,
      "instructionfollowing": 15221,
      "instructgpt": 15136,
      "simulator": 30191,
      "session": 29667,
      "automated": 2854,
      "inline": 14985,
      "multiturn": 22460,
      "interactions": 15387,
      "asks": 2560,
      "simulated": 30185,
      "studying": 31410,
      "gpt2based": 12972,
      "singleturn": 30235,
      "mixed": 20269,
      "codeswitching": 4858,
      "conversation": 6089,
      "phenomenon": 24922,
      "prominent": 26264,
      "media": 19758,
      "platforms": 25010,
      "codemixed": 4843,
      "gaining": 12067,
      "popularity": 25140,
      "codemixing": 4845,
      "roman": 28965,
      "twitter": 34037,
      "modelling": 20913,
      "pos": 25154,
      "tweets": 34035,
      "capturing": 4155,
      "outlined": 23743,
      "returns": 28783,
      "dictionaries": 8028,
      "patients": 24405,
      "provider": 26744,
      "indian": 14682,
      "tackle": 31994,
      "faced": 10826,
      "berts": 3545,
      "mlm": 20301,
      "inspection": 15088,
      "intervention": 15476,
      "spurred": 30837,
      "interpreting": 15467,
      "executing": 10196,
      "behavioral": 3324,
      "analyzing": 2001,
      "salience": 29059,
      "backbone": 3052,
      "interprets": 15469,
      "debugging": 7217,
      "disambiguation": 8252,
      "identifying": 14019,
      "interventions": 15477,
      "tailor": 32014,
      "refers": 27777,
      "satisfy": 29110,
      "attributes": 2762,
      "emotions": 9207,
      "resort": 28405,
      "plm": 25036,
      "concatenated": 5520,
      "decrease": 7296,
      "position": 25180,
      "sensitivity": 29520,
      "bridge": 3863,
      "concatenating": 5522,
      "connector": 5688,
      "attributespecific": 2765,
      "styles": 31416,
      "davinci": 7199,
      "yields": 35920,
      "moderate": 22150,
      "textdavinci002": 32979,
      "grounded": 13287,
      "flawed": 11613,
      "cooccur": 6126,
      "observations": 23221,
      "reliability": 27947,
      "idioms": 14030,
      "figurative": 11159,
      "cultures": 6462,
      "pose": 25156,
      "mt": 22250,
      "idiomatic": 14029,
      "macro": 19376,
      "dialogpt": 8006,
      "idiom": 14028,
      "hub": 13776,
      "knows": 15938,
      "resolution": 28400,
      "annotate": 2014,
      "qabased": 26919,
      "discern": 8256,
      "return": 28781,
      "valid": 34988,
      "mentions": 19845,
      "inconsistent": 14540,
      "polish": 25092,
      "initializing": 14970,
      "plbart": 25034,
      "prove": 26667,
      "equivalent": 9687,
      "revisiting": 28839,
      "arabic": 2425,
      "body": 3803,
      "addressing": 1368,
      "wellexplored": 35524,
      "methodical": 19987,
      "revisit": 28838,
      "perspectives": 24899,
      "discriminative": 8285,
      "shortly": 29826,
      "life": 17821,
      "overlooking": 23948,
      "variability": 35031,
      "psycholinguistic": 26820,
      "albert": 1695,
      "fashion": 10989,
      "skill": 30307,
      "decisions": 7240,
      "predictive": 25447,
      "framed": 11821,
      "recognizing": 27646,
      "entailment": 9615,
      "aka": 1676,
      "nli": 22911,
      "classical": 4589,
      "spurious": 30834,
      "exists": 10330,
      "harder": 13425,
      "expressions": 10643,
      "spanning": 30596,
      "simile": 30138,
      "modelintheloop": 20912,
      "crowd": 6429,
      "workers": 35800,
      "annotators": 2034,
      "conjunction": 5678,
      "aid": 1626,
      "bring": 3877,
      "typing": 34083,
      "emotion": 9201,
      "cardinality": 4163,
      "informative": 14931,
      "orders": 23680,
      "combinatorial": 4954,
      "taking": 32035,
      "gets": 12733,
      "days": 7202,
      "seconds": 29337,
      "exams": 10145,
      "exam": 10094,
      "institution": 15127,
      "mit": 20246,
      "harvard": 13468,
      "cornell": 6160,
      "write": 35845,
      "students": 31258,
      "hours": 13766,
      "finals": 11207,
      "level": 17729,
      "program": 26193,
      "synthesis": 31835,
      "stem": 31033,
      "differ": 8033,
      "broader": 3890,
      "curate": 6466,
      "course": 6310,
      "notes": 23036,
      "checkers": 4516,
      "numeric": 23176,
      "chainofthought": 4295,
      "highlight": 13626,
      "transformative": 33696,
      "streamline": 31134,
      "workload": 35807,
      "mere": 19846,
      "harness": 13452,
      "completeness": 5255,
      "originality": 23725,
      "thinking": 33067,
      "availability": 2962,
      "nonautoregressive": 22988,
      "dominated": 8662,
      "yield": 35907,
      "decoder": 7246,
      "inefficient": 14750,
      "singlestep": 30233,
      "inferior": 14830,
      "variables": 35037,
      "interdependence": 15401,
      "termed": 32732,
      "predictor": 25456,
      "glancing": 12786,
      "sampler": 29070,
      "decoders": 7266,
      "rate": 27260,
      "aishell1": 1675,
      "20000": 218,
      "hour": 13764,
      "attain": 2695,
      "10x": 65,
      "allinone": 1797,
      "taskindependent": 32216,
      "heuristic": 13531,
      "trivial": 33919,
      "lowquality": 19298,
      "combat": 4947,
      "condense": 5562,
      "inherent": 14945,
      "reformulates": 27801,
      "granularity": 13216,
      "deberta": 7213,
      "conll03": 5680,
      "transfers": 33692,
      "try": 33945,
      "decipher": 7230,
      "connection": 5684,
      "decades": 7219,
      "essence": 9755,
      "viewed": 35274,
      "accessing": 835,
      "principle": 25931,
      "consist": 5729,
      "valuable": 35010,
      "overcoming": 23927,
      "competitors": 5236,
      "college": 4939,
      "entrance": 9658,
      "examination": 10096,
      "authoritative": 2836,
      "china": 4539,
      "116": 80,
      "mark": 19591,
      "150": 145,
      "gaokao": 12085,
      "2022": 229,
      "happened": 13418,
      "ago": 1588,
      "108": 62,
      "showcase": 29834,
      "77": 520,
      "stimulate": 31062,
      "generality": 12199,
      "93": 597,
      "flant5": 11591,
      "character": 4421,
      "cultural": 6458,
      "entertainment": 9621,
      "occasionally": 23268,
      "supplemented": 31703,
      "pronunciation": 26458,
      "simplified": 30177,
      "characters": 4436,
      "produces": 26159,
      "filtered": 11171,
      "retrievalbased": 28765,
      "chatglm": 4454,
      "stored": 31081,
      "carried": 4186,
      "presenting": 25570,
      "acquired": 1120,
      "stages": 30860,
      "levels": 17738,
      "morphology": 22223,
      "inconsistently": 14541,
      "perceived": 24456,
      "induced": 14732,
      "endeavors": 9424,
      "sector": 29341,
      "clear": 4646,
      "earlier": 8770,
      "maintained": 19417,
      "infer": 14757,
      "albeit": 1693,
      "inductive": 14737,
      "compositionality": 5329,
      "encodes": 9380,
      "symbols": 31804,
      "correspond": 6222,
      "nodes": 22982,
      "infers": 14833,
      "posterior": 25225,
      "uncover": 34134,
      "groundtruth": 13293,
      "inferred": 14831,
      "sentiments": 29573,
      "symbolic": 31800,
      "walk": 35409,
      "described": 7677,
      "corpusbased": 6189,
      "dilemma": 8193,
      "free": 11909,
      "dependence": 7615,
      "aforementioned": 1551,
      "detailed": 7830,
      "lemmatization": 17701,
      "grouping": 13301,
      "forms": 11766,
      "analysed": 1906,
      "item": 15679,
      "identified": 13998,
      "linguistics": 18026,
      "algorithmic": 1716,
      "determining": 7902,
      "intended": 15367,
      "stemming": 31034,
      "depends": 7622,
      "surrounding": 31770,
      "google": 12825,
      "run": 29012,
      "lengths": 17714,
      "bank": 3093,
      "remember": 28071,
      "regards": 27813,
      "keyvalue": 15793,
      "knowledgeable": 15924,
      "slots": 30330,
      "interpretable": 15460,
      "salient": 29060,
      "fix": 11576,
      "sure": 31722,
      "influenced": 14843,
      "mounting": 22234,
      "closedbook": 4682,
      "degrade": 7378,
      "interpretability": 15456,
      "keys": 15791,
      "humanreadable": 13915,
      "influence": 14837,
      "cognitive": 4873,
      "powered": 25331,
      "pervasive": 24914,
      "day": 7201,
      "decisionmaking": 7235,
      "engaging": 9456,
      "shed": 29793,
      "recruited": 27673,
      "amateur": 1867,
      "positively": 25198,
      "negatively": 22664,
      "opinions": 23579,
      "align": 1724,
      "interact": 15379,
      "incorporated": 14570,
      "criteria": 6376,
      "usual": 34943,
      "plans": 25004,
      "lens": 17718,
      "theoretical": 33047,
      "movie": 22238,
      "review": 28827,
      "followed": 11682,
      "speak": 30647,
      "initiation": 14975,
      "initiate": 14973,
      "turns": 34031,
      "period": 24859,
      "realtime": 27330,
      "feedback": 11056,
      "flow": 11627,
      "respond": 28466,
      "prosodic": 26653,
      "operating": 23564,
      "audio": 2774,
      "transcriptions": 33665,
      "wrt": 35866,
      "true": 33923,
      "switchboard": 31797,
      "waiting": 35407,
      "momentum": 22205,
      "optimizers": 23652,
      "chosen": 4562,
      "trials": 33893,
      "acceleration": 803,
      "avoids": 3040,
      "adopts": 1414,
      "secondorder": 29335,
      "moments": 22204,
      "finds": 11262,
      "firstorder": 11567,
      "stationary": 31016,
      "matching": 19658,
      "bound": 3842,
      "rl": 28898,
      "sotas": 30540,
      "resnet": 28398,
      "convnext": 6120,
      "swin": 31794,
      "mae": 19379,
      "half": 13365,
      "epochs": 9672,
      "tolerance": 33253,
      "1k": 202,
      "32k": 325,
      "phishing": 24931,
      "psychological": 26822,
      "trait": 33656,
      "traits": 33657,
      "legitimate": 17700,
      "urgency": 34494,
      "fear": 11015,
      "desire": 7760,
      "nuances": 23134,
      "concatenate": 5519,
      "imbalanced": 14103,
      "strongest": 31204,
      "cue": 6455,
      "sparql": 30607,
      "triples": 33914,
      "express": 10637,
      "aggregation": 1580,
      "urgently": 34497,
      "forward": 11781,
      "handle": 13406,
      "rephrase": 28096,
      "nl": 22905,
      "smoothing": 30411,
      "factoid": 10862,
      "collaborative": 4906,
      "bloom176b": 3791,
      "opt175b": 23605,
      "download": 8669,
      "highend": 13592,
      "unavailable": 34112,
      "offloading": 23324,
      "innate": 14986,
      "logits": 19162,
      "collaboratively": 4908,
      "parties": 24358,
      "running": 29014,
      "consumer": 5817,
      "approx": 2410,
      "natively": 22501,
      "exposes": 10633,
      "served": 29651,
      "share": 29780,
      "custom": 6556,
      "triplets": 33917,
      "shared": 29783,
      "submissions": 31431,
      "identification": 13996,
      "casual": 4211,
      "spans": 30601,
      "iteratively": 15689,
      "conditioning": 5577,
      "triplet": 33916,
      "component": 5308,
      "160": 158,
      "placed": 24984,
      "assuming": 2662,
      "readability": 27294,
      "ensembles": 9598,
      "reliable": 27953,
      "ranging": 27216,
      "reliably": 27956,
      "ensemble": 9596,
      "germeval": 12732,
      "outofsample": 23756,
      "root": 28969,
      "mean": 19717,
      "gradientbased": 13192,
      "trends": 33889,
      "everlarger": 10053,
      "hyperparameter": 13958,
      "offers": 23305,
      "tune": 33955,
      "bayesian": 3295,
      "schedules": 29228,
      "concurrently": 5561,
      "schedule": 29226,
      "explainable": 10530,
      "greedy": 13280,
      "multidomain": 22279,
      "attempts": 2708,
      "summary": 31632,
      "project": 26243,
      "python": 26902,
      "gptstyle": 13165,
      "eval": 9815,
      "dfx": 7997,
      "lowlatency": 19296,
      "accelerating": 798,
      "services": 29659,
      "datacenters": 6921,
      "degrades": 7382,
      "characteristic": 4424,
      "executes": 10195,
      "parallelism": 24173,
      "dataflow": 6929,
      "simultaneous": 30193,
      "execution": 10197,
      "cores": 6159,
      "operate": 23561,
      "xilinx": 35870,
      "alveo": 1860,
      "u280": 34087,
      "fpgas": 11814,
      "channels": 4419,
      "bandwidth": 3088,
      "hbm": 13476,
      "v100": 34981,
      "workloads": 35808,
      "cloud": 4702,
      "array": 2509,
      "proliferation": 26257,
      "highstakes": 13711,
      "medicine": 19771,
      "burgeoning": 3943,
      "transparency": 33861,
      "greater": 13266,
      "1000x": 47,
      "instantiations": 15118,
      "naturallanguage": 22583,
      "augments": 2829,
      "decoupled": 7293,
      "expansions": 10346,
      "counterparts": 6302,
      "6billion": 490,
      "gptj": 13145,
      "transparent": 33863,
      "fmri": 11640,
      "interpretations": 15464,
      "scientific": 29247,
      "reproducing": 28204,
      "github": 12738,
      "evidence": 10059,
      "judgments": 15727,
      "typical": 34075,
      "direct": 8209,
      "unusual": 34465,
      "naively": 22479,
      "memorise": 19792,
      "statistics": 31023,
      "repeatedly": 28091,
      "continue": 5985,
      "scaled": 29149,
      "palm": 23988,
      "objects": 23214,
      "closely": 4692,
      "overfitting": 23932,
      "characterized": 4430,
      "cooccurrences": 6129,
      "intrinsic": 15489,
      "extrinsic": 10809,
      "science": 29240,
      "metadata": 19860,
      "abstracts": 778,
      "compiled": 5239,
      "obtained": 23255,
      "sparsely": 30622,
      "keywords": 15797,
      "labelling": 15965,
      "stories": 31084,
      "transcripts": 33666,
      "outofthebox": 23757,
      "brittle": 3882,
      "designing": 7751,
      "producing": 26164,
      "imperfect": 14154,
      "aggregating": 1579,
      "motivate": 22226,
      "ama": 1865,
      "went": 35533,
      "restrict": 28530,
      "recursively": 27688,
      "votes": 35391,
      "accuracies": 856,
      "lift": 17824,
      "102": 55,
      "gptj6b": 13152,
      "match": 19639,
      "exceed": 10146,
      "gpt3175b": 13011,
      "averaged": 3027,
      "html": 13772,
      "exceptional": 10166,
      "webpage": 35483,
      "automation": 2925,
      "webbased": 35480,
      "autonomous": 2928,
      "navigation": 22596,
      "remarkably": 28065,
      "exclusively": 10187,
      "miniwob": 20203,
      "distilled": 8349,
      "decaying": 7221,
      "pertoken": 24907,
      "kernelbased": 15750,
      "retains": 28720,
      "99": 610,
      "attentions": 2750,
      "substitutes": 31491,
      "sports": 30819,
      "disambiguate": 8251,
      "fusion": 12021,
      "amenable": 1877,
      "solved": 30501,
      "offtheshelf": 23326,
      "optional": 23661,
      "possibly": 25219,
      "fuse": 12017,
      "paragraph": 24162,
      "dart": 6584,
      "old": 23334,
      "feasibility": 11016,
      "failures": 10920,
      "failure": 10916,
      "action": 1130,
      "feasible": 11019,
      "binary": 3731,
      "multichoice": 22271,
      "mcq": 19715,
      "19": 191,
      "62": 464,
      "64": 472,
      "instructionfinetuned": 15217,
      "phrased": 24937,
      "dramatically": 8717,
      "upalm": 34471,
      "setups": 29750,
      "cot": 6279,
      "mmlu": 20312,
      "bbh": 3296,
      "mgsm": 20153,
      "flanpalm": 11589,
      "540b": 443,
      "18k": 190,
      "checkpoints": 4521,
      "62b": 467,
      "usability": 34499,
      "mandarin": 19533,
      "grouped": 13298,
      "highlevel": 13622,
      "lost": 19258,
      "assign": 2622,
      "acceptable": 811,
      "contains": 5838,
      "transformations": 33695,
      "90": 589,
      "severe": 29752,
      "18": 183,
      "xlm": 35874,
      "highest": 13618,
      "gender": 12148,
      "communicate": 5046,
      "completely": 5254,
      "refer": 27770,
      "independently": 14671,
      "conclusion": 5556,
      "matches": 19649,
      "indistribution": 14713,
      "13b": 117,
      "crawl": 6344,
      "removing": 28076,
      "barriers": 3103,
      "emergence": 9166,
      "100b": 48,
      "notably": 23028,
      "emergent": 9181,
      "sheer": 29801,
      "openaccess": 23428,
      "billionparameter": 3723,
      "practices": 25379,
      "englishonly": 9499,
      "shape": 29775,
      "semiparametric": 29498,
      "fullyparametric": 11959,
      "zerofewshot": 35947,
      "evolving": 10080,
      "empowers": 9272,
      "parametric": 24296,
      "retrieves": 28779,
      "fed": 11047,
      "interestingly": 15412,
      "selector": 29400,
      "plays": 25025,
      "determine": 7896,
      "assignment": 2627,
      "inspires": 15102,
      "770m": 521,
      "gained": 12059,
      "stems": 31035,
      "overwhelming": 23957,
      "term": 32729,
      "semiconductor": 29494,
      "rouge": 28977,
      "exactly": 10092,
      "judgment": 15726,
      "estimating": 9784,
      "carbon": 4158,
      "footprint": 11712,
      "176b": 178,
      "quantify": 26986,
      "cycle": 6574,
      "blooms": 3793,
      "approximately": 2413,
      "247": 264,
      "consumption": 5823,
      "account": 850,
      "equipment": 9679,
      "manufacturing": 19573,
      "operational": 23566,
      "emissions": 9199,
      "endpoint": 9429,
      "multihop": 22286,
      "decompose": 7284,
      "decomposer": 7286,
      "conciseness": 5549,
      "overlooked": 23946,
      "define": 7362,
      "simplification": 30171,
      "2000": 217,
      "organizations": 23691,
      "frequently": 11925,
      "kept": 15748,
      "democratizing": 7426,
      "collaboration": 4904,
      "comprising": 5433,
      "46": 397,
      "59": 453,
      "breakthroughs": 3862,
      "owing": 23958,
      "internalize": 15444,
      "interacts": 15397,
      "taskrelevant": 32224,
      "conflicts": 5666,
      "memorized": 19795,
      "ignore": 14037,
      "undertake": 34284,
      "aware": 3041,
      "strengthen": 31139,
      "showcases": 29845,
      "look": 19219,
      "ood": 23378,
      "evolves": 10079,
      "codegen": 4838,
      "scan": 29183,
      "exemplars": 10204,
      "decreasing": 7302,
      "galactica": 12075,
      "overload": 23943,
      "obstacle": 23243,
      "explosive": 10625,
      "growth": 13321,
      "mass": 19621,
      "today": 33182,
      "engines": 9473,
      "unable": 34100,
      "organize": 23692,
      "material": 19662,
      "technical": 32604,
      "probes": 25978,
      "equations": 9677,
      "latest": 17416,
      "682": 486,
      "versus": 35238,
      "357": 350,
      "math": 19666,
      "204": 242,
      "88": 581,
      "pubmedqa": 26875,
      "medmcqa": 19778,
      "dev": 7906,
      "bigbench": 3700,
      "links": 18031,
      "transition": 33808,
      "f1score": 10819,
      "conll": 5679,
      "41": 383,
      "743": 513,
      "53": 439,
      "f1scores": 10820,
      "backpropagation": 3067,
      "prefixes": 25483,
      "lora": 19225,
      "modified": 22180,
      "trains": 33653,
      "mtf": 22262,
      "landscape": 16022,
      "leaderboards": 17472,
      "variant": 35040,
      "analogies": 1902,
      "examines": 10106,
      "analogy": 1904,
      "analogical": 1901,
      "holds": 13740,
      "dissimilar": 8325,
      "posed": 25163,
      "largelanguage": 17303,
      "llama2": 18158,
      "mpt": 22245,
      "falcon": 10937,
      "attentionhead": 2749,
      "visualization": 35360,
      "impossible": 14225,
      "unlikely": 34408,
      "surprising": 31764,
      "understudied": 34282,
      "acquire": 1114,
      "minimally": 20192,
      "implausible": 14155,
      "possess": 25201,
      "teacher": 32581,
      "laptop": 16925,
      "preferences": 25471,
      "followup": 11708,
      "passive": 24386,
      "synonymous": 31811,
      "mirror": 20207,
      "iv": 15694,
      "organizing": 23694,
      "dimension": 8196,
      "documents": 8516,
      "dominate": 8661,
      "directed": 8219,
      "fine": 11263,
      "hotpotqa": 13763,
      "chunks": 4564,
      "chunk": 4563,
      "mbert": 19710,
      "xlmr": 35875,
      "mbart": 19709,
      "highresource": 13705,
      "unannotated": 34106,
      "appears": 2115,
      "connections": 5686,
      "modeled": 20881,
      "interpret": 15453,
      "favorably": 11011,
      "logarithmic": 19153,
      "planning": 24999,
      "obtaining": 23261,
      "automata": 2849,
      "glms": 12793,
      "formally": 11751,
      "constructs": 5813,
      "finite": 11561,
      "automaton": 2926,
      "brief": 3875,
      "builds": 3928,
      "textbased": 32971,
      "fills": 11168,
      "userdefined": 34678,
      "specifications": 30763,
      "accordingly": 848,
      "refine": 27779,
      "outcomes": 23735,
      "verification": 35209,
      "crossing": 6410,
      "highlyspecialized": 13672,
      "secure": 29344,
      "multiparty": 22374,
      "finnish": 11562,
      "journalists": 15717,
      "house": 13768,
      "advances": 1477,
      "price": 25905,
      "formidable": 11763,
      "proportional": 26483,
      "rapidly": 27254,
      "realize": 27320,
      "convenient": 6070,
      "layerwise": 17450,
      "dropping": 8745,
      "125x": 99,
      "smallscale": 30405,
      "pronouns": 26457,
      "subject": 31423,
      "object": 23198,
      "corresponds": 6233,
      "verb": 35202,
      "frequency": 11923,
      "verbs": 35207,
      "raters": 27269,
      "explain": 10523,
      "easier": 8788,
      "literacy": 18037,
      "involves": 15627,
      "rephrasing": 28098,
      "purely": 26881,
      "sari": 29101,
      "mechanical": 19745,
      "turk": 34027,
      "bigscience": 3705,
      "workshop": 35831,
      "initiative": 14976,
      "interdisciplinary": 15403,
      "collaborations": 4905,
      "ethics": 9808,
      "governance": 12837,
      "participant": 24324,
      "carry": 4190,
      "lessons": 17722,
      "did": 8030,
      "goes": 12811,
      "basis": 3284,
      "inception": 14443,
      "reused": 28786,
      "activated": 1135,
      "decouple": 7292,
      "attractive": 2754,
      "datahungry": 6931,
      "checkpoint": 4518,
      "initial": 14958,
      "judgements": 15723,
      "unacceptable": 34102,
      "contextfree": 5932,
      "contextualized": 5960,
      "mismatch": 20220,
      "stability": 30846,
      "grammaticality": 13210,
      "randomly": 27180,
      "unstable": 34449,
      "worsen": 35839,
      "violated": 35278,
      "amplified": 1894,
      "overlap": 23940,
      "explained": 10531,
      "66": 479,
      "hypothesis": 13963,
      "uniformly": 34345,
      "spread": 30827,
      "opt66b": 23609,
      "heads": 13480,
      "removed": 28075,
      "decline": 7243,
      "unimportant": 34350,
      "primitive": 25928,
      "reinforcing": 27848,
      "arguments": 2494,
      "head": 13477,
      "quantities": 27000,
      "cater": 4232,
      "casting": 4209,
      "eliminating": 9109,
      "unnatural": 34416,
      "labor": 15970,
      "inferencetime": 14826,
      "vast": 35180,
      "virtually": 35285,
      "eliciting": 9105,
      "fourth": 11813,
      "fair": 10922,
      "rivals": 28897,
      "diversification": 8473,
      "personality": 24884,
      "psychology": 26823,
      "diagnose": 7998,
      "exhibited": 10233,
      "asked": 2555,
      "party": 24370,
      "diagnostic": 8001,
      "markers": 19597,
      "illustrates": 14049,
      "manipulated": 19540,
      "predictable": 25412,
      "frames": 11822,
      "personas": 24892,
      "subjects": 31427,
      "collated": 4910,
      "reddit": 27692,
      "interleaving": 15420,
      "promptingbased": 26400,
      "uptodate": 34489,
      "onestep": 23354,
      "retrieveandread": 28770,
      "depend": 7613,
      "retrieved": 28771,
      "interleaves": 15419,
      "musique": 22470,
      "flant5large": 11603,
      "hallucination": 13372,
      "safety": 29041,
      "unbiased": 34113,
      "inventory": 15566,
      "scored": 29275,
      "gpt35": 13012,
      "gpt4": 13049,
      "llama2chat7b": 18222,
      "preference": 25466,
      "recommended": 27654,
      "projection": 26249,
      "readily": 27300,
      "formulated": 11773,
      "pertaining": 24904,
      "decomposes": 7287,
      "candidates": 3994,
      "ranked": 27226,
      "african": 1554,
      "sparsegpt": 30621,
      "pruned": 26803,
      "gptfamily": 13142,
      "45": 393,
      "reach": 27286,
      "negligible": 22670,
      "ignored": 14038,
      "semistructured": 29502,
      "diffusion": 8179,
      "concretely": 5559,
      "revolutionizing": 28853,
      "sectors": 29342,
      "job": 15707,
      "positions": 25190,
      "transformed": 33698,
      "dalle2": 6581,
      "3d": 368,
      "flamingo": 11585,
      "video": 35262,
      "taxonomy": 32574,
      "discoveries": 8271,
      "provable": 26665,
      "maximal": 19699,
      "regularizer": 27828,
      "mmr": 20316,
      "multihead": 22284,
      "self": 29403,
      "corroborate": 6234,
      "infusion": 14942,
      "adopting": 1406,
      "usercentric": 34677,
      "chatbots": 4451,
      "volumes": 35388,
      "limiting": 17975,
      "stylistic": 31417,
      "memorability": 19790,
      "empathy": 9208,
      "infusing": 14941,
      "comparisons": 5201,
      "bootstrap": 3830,
      "infuse": 14939,
      "balancing": 3086,
      "compelling": 5207,
      "stylized": 31418,
      "realm": 27327,
      "contributes": 6032,
      "affect": 1536,
      "nm": 22976,
      "masks": 19620,
      "innovations": 14992,
      "a100": 613,
      "nontrivial": 23013,
      "attentionbased": 2746,
      "origins": 23730,
      "poorly": 25106,
      "estimated": 9782,
      "moment": 22203,
      "variance": 35038,
      "phases": 24919,
      "calculates": 3963,
      "trajectory": 33660,
      "concentration": 5525,
      "mitigates": 20261,
      "aggressive": 1581,
      "ratios": 27280,
      "tabular": 31991,
      "independent": 14670,
      "parent": 24308,
      "table": 31983,
      "tables": 31986,
      "captures": 4154,
      "needing": 22648,
      "retrievers": 28778,
      "plugin": 25057,
      "option": 23660,
      "instantiate": 15115,
      "included": 14448,
      "beir": 3335,
      "necessity": 22614,
      "procedural": 26042,
      "straightforward": 31091,
      "reflects": 27798,
      "intentions": 15377,
      "incredibly": 14652,
      "textprompted": 32985,
      "dynamics": 8768,
      "discovery": 8274,
      "varieties": 35055,
      "weaknesses": 35464,
      "revealed": 28808,
      "said": 29057,
      "connect": 5681,
      "expect": 10348,
      "polysemy": 25098,
      "spaces": 30586,
      "street": 31137,
      "premises": 25494,
      "conclusions": 5557,
      "enabled": 9290,
      "chat": 4439,
      "lists": 18035,
      "lamda": 16021,
      "aspectbased": 2569,
      "positive": 25192,
      "neutral": 22771,
      "16": 156,
      "ate": 2678,
      "337": 331,
      "experiences": 10374,
      "misleading": 20218,
      "hybrid": 13954,
      "cqa": 6338,
      "cps": 6334,
      "want": 35419,
      "freedom": 11911,
      "mix": 20268,
      "climate": 4654,
      "protection": 26657,
      "explainability": 10525,
      "versatile": 35223,
      "allowed": 1803,
      "nonspecialists": 23009,
      "start": 30897,
      "edited": 8824,
      "helm": 13500,
      "strict": 31145,
      "multimodal": 22336,
      "neurosymbolic": 22770,
      "directional": 8226,
      "stimulus": 31065,
      "act": 1128,
      "nuanced": 23132,
      "sidesteps": 29940,
      "offline": 23320,
      "multiwoz": 22467,
      "enhances": 9545,
      "chatgpts": 4506,
      "humancrafted": 13894,
      "lags": 16016,
      "weighted": 35499,
      "rwkv": 29022,
      "45m": 396,
      "rendering": 28080,
      "quadratic": 26926,
      "20x": 248,
      "democratize": 7424,
      "locate": 19144,
      "cuttingedge": 6570,
      "attribution": 2768,
      "drive": 8734,
      "llama": 18058,
      "7b": 524,
      "65b": 476,
      "trillions": 33909,
      "resorting": 28406,
      "proprietary": 26634,
      "inaccessible": 14431,
      "llama13b": 18156,
      "llama65b": 18228,
      "palm540b": 23997,
      "accompanied": 838,
      "commensurate": 4979,
      "continuation": 5983,
      "engine": 9458,
      "offering": 23296,
      "fuzzy": 12053,
      "exact": 10087,
      "hugging": 13783,
      "face": 10821,
      "functional": 11964,
      "neglecting": 22668,
      "longstanding": 19213,
      "considered": 5716,
      "illustrative": 14053,
      "summarize": 31628,
      "translate": 33811,
      "matter": 19695,
      "dont": 8663,
      "humanlevel": 13903,
      "generalizability": 12200,
      "push": 26892,
      "collections": 4935,
      "varied": 35050,
      "british": 3881,
      "american": 1878,
      "conventions": 6081,
      "establish": 9765,
      "appear": 2112,
      "correcting": 6197,
      "debiased": 7214,
      "somewhat": 30520,
      "composite": 5324,
      "pressing": 25615,
      "international": 15446,
      "formed": 11762,
      "undertaking": 34286,
      "putting": 26897,
      "harm": 13437,
      "curation": 6474,
      "undertaken": 34285,
      "thereof": 33062,
      "empower": 9265,
      "lossless": 19254,
      "expectation": 10349,
      "prunes": 26807,
      "segmentation": 29370,
      "imagenet1k": 14080,
      "save": 29113,
      "248": 265,
      "illustrating": 14050,
      "concern": 5538,
      "societies": 30439,
      "completing": 5257,
      "130": 110,
      "co2e": 4714,
      "doing": 8546,
      "substitute": 31490,
      "activities": 1151,
      "kbqa": 15743,
      "supports": 31717,
      "follows": 11707,
      "kbbased": 15742,
      "commonalities": 5021,
      "resemble": 28383,
      "vicuna": 35246,
      "comprehending": 5337,
      "black": 3751,
      "devised": 7993,
      "comprehend": 5334,
      "sounds": 30544,
      "iterations": 15682,
      "meanings": 19728,
      "continued": 5989,
      "identity": 14026,
      "corrupted": 6236,
      "reasonable": 27359,
      "fallacies": 10952,
      "preferred": 25476,
      "convey": 6116,
      "lesser": 17720,
      "targets": 32068,
      "cameras": 3989,
      "opinion": 23577,
      "generationbased": 12640,
      "extractor": 10781,
      "modelname": 20916,
      "assistant": 2636,
      "ubiquitous": 34088,
      "revisions": 28837,
      "pronounced": 26456,
      "bangla": 3092,
      "grammatical": 13204,
      "symbol": 31799,
      "postprocessing": 25228,
      "distance": 8326,
      "detected": 7853,
      "grammar": 13203,
      "replicate": 28108,
      "instructiontuning": 15301,
      "factor": 10863,
      "merely": 19847,
      "curve": 6553,
      "koala": 15939,
      "index": 14680,
      "overlaps": 23942,
      "suffix": 31565,
      "proportion": 26482,
      "forensic": 11726,
      "visual": 35325,
      "revolutionized": 28843,
      "unfortunately": 34312,
      "incorrect": 14581,
      "renders": 28081,
      "subfields": 31421,
      "integers": 15317,
      "judging": 15725,
      "integer": 15316,
      "divided": 8489,
      "prime": 25924,
      "computers": 5509,
      "classifying": 4632,
      "critically": 6398,
      "automl": 2927,
      "failed": 10909,
      "engineered": 9460,
      "era": 9691,
      "chatdoctor": 4453,
      "alpaca": 1821,
      "peft": 24434,
      "easytouse": 8803,
      "adapters": 1204,
      "placement": 24985,
      "locations": 19149,
      "fourteen": 11812,
      "arithmetic": 2499,
      "smallerscale": 30402,
      "chef": 4525,
      "imagery": 14081,
      "intelligent": 15362,
      "familiar": 10963,
      "captioning": 4135,
      "monolithic": 22211,
      "cards": 4164,
      "dietary": 8032,
      "restrictions": 28535,
      "concludes": 5554,
      "struggled": 31248,
      "repetitive": 28094,
      "cook": 6130,
      "pragmatic": 25383,
      "featuring": 11046,
      "parrot": 24313,
      "processingnlp": 26135,
      "accomplished": 843,
      "restricted": 28531,
      "regulate": 27830,
      "wmt22": 35631,
      "details": 7844,
      "cerebrasgpt": 4270,
      "computeoptimal": 5501,
      "cluster": 4706,
      "111m": 76,
      "powerlaw": 25355,
      "learnings": 17676,
      "parameterization": 24217,
      "predictability": 25411,
      "aiding": 1628,
      "expressed": 10639,
      "norm": 23015,
      "structureaware": 31217,
      "universally": 34373,
      "uie": 34091,
      "linearized": 17998,
      "unleashing": 34390,
      "induce": 14730,
      "structural": 31209,
      "posttraining": 25232,
      "compact": 5064,
      "trees": 33880,
      "highorder": 13673,
      "helping": 13521,
      "endtasks": 9431,
      "taskadaptive": 32211,
      "resolves": 28403,
      "longrange": 19211,
      "boundary": 3844,
      "education": 8834,
      "teachers": 32587,
      "alike": 1787,
      "judge": 15720,
      "sufficiently": 31563,
      "classroom": 4633,
      "unhelpful": 34317,
      "usefulness": 34643,
      "taxonomies": 32573,
      "ts": 33949,
      "word2vec": 35652,
      "userfriendly": 34679,
      "44": 392,
      "lowcost": 19277,
      "akin": 1677,
      "fostering": 11787,
      "quantity": 27001,
      "encompassing": 9389,
      "supplement": 31701,
      "advancement": 1446,
      "closest": 4699,
      "secondary": 29330,
      "34b": 338,
      "fullparameter": 11948,
      "lorabased": 19236,
      "foundational": 11801,
      "reproduction": 28205,
      "mitigating": 20262,
      "educators": 8848,
      "evolutionary": 10076,
      "taught": 32572,
      "university": 34376,
      "south": 30577,
      "strides": 31147,
      "agi": 1583,
      "nonetheless": 22994,
      "obstacles": 23244,
      "predominantly": 25461,
      "llamas": 18241,
      "markedly": 19596,
      "proficiency": 26178,
      "ceval": 4286,
      "dataefficient": 6927,
      "evergrowing": 10050,
      "equipped": 9680,
      "budgets": 3902,
      "homogeneous": 13751,
      "pretrains": 25853,
      "threestep": 33095,
      "kmeans": 15803,
      "suitability": 31595,
      "longform": 19204,
      "alignment": 1751,
      "cheaper": 4510,
      "instructiontuned": 15283,
      "things": 33064,
      "premise": 25493,
      "basic": 3282,
      "functioning": 11969,
      "urgent": 34495,
      "representational": 28153,
      "articulate": 2526,
      "pertains": 24905,
      "exercise": 10207,
      "broadly": 3893,
      "ordinary": 23685,
      "demystifying": 7599,
      "mystery": 22476,
      "informed": 14935,
      "consensus": 5693,
      "expansive": 10347,
      "whos": 35543,
      "mls": 20308,
      "incoherent": 14534,
      "grade": 13183,
      "longterm": 19216,
      "immediately": 14112,
      "automate": 2850,
      "shots": 29833,
      "reside": 28387,
      "recursive": 27687,
      "faces": 10827,
      "questionanswer": 27078,
      "display": 8314,
      "stanford": 30891,
      "massively": 19634,
      "27b": 284,
      "67b": 484,
      "performant": 24824,
      "usd": 34511,
      "slot": 30328,
      "filling": 11166,
      "participation": 24333,
      "2023": 232,
      "mt0": 22252,
      "intentionally": 15375,
      "humanlabeled": 13902,
      "llmaugmented": 18383,
      "navigate": 22592,
      "timeintensive": 33152,
      "acquiring": 1123,
      "annotating": 2023,
      "guidelines": 13356,
      "synthetically": 31863,
      "proves": 26679,
      "rare": 27258,
      "multiclass": 22272,
      "moderately": 22151,
      "exponentially": 10628,
      "percentage": 24459,
      "uniform": 34344,
      "choosing": 4560,
      "contrary": 6007,
      "slower": 30333,
      "posit": 25179,
      "selectively": 29399,
      "distilling": 8356,
      "reaction": 27291,
      "llmgenerated": 18393,
      "needed": 22646,
      "extracts": 10782,
      "rationales": 27278,
      "tracking": 33327,
      "prerequisite": 25506,
      "trainingevaluation": 33648,
      "taken": 32022,
      "conformal": 5668,
      "nucleus": 23135,
      "successively": 31546,
      "chooses": 4559,
      "cumulative": 6465,
      "exceeds": 10151,
      "entropy": 9661,
      "overconfident": 23930,
      "inverse": 15567,
      "authors": 2837,
      "suddenly": 31548,
      "expertise": 10513,
      "myriad": 22474,
      "evolve": 10077,
      "probably": 25975,
      "rapid": 27237,
      "ushering": 34719,
      "profound": 26191,
      "govern": 12836,
      "wisely": 35622,
      "poses": 25165,
      "industries": 14743,
      "maximize": 19702,
      "minimize": 20195,
      "reparameterization": 28088,
      "constitute": 5775,
      "reaches": 27287,
      "hurting": 13952,
      "inferring": 14832,
      "tag": 32009,
      "treated": 33872,
      "lieu": 17820,
      "near": 22598,
      "roughly": 28985,
      "supervising": 31694,
      "bits": 3750,
      "clustering": 4708,
      "8bit": 583,
      "threefold": 33093,
      "71": 507,
      "32gb": 324,
      "sts": 31251,
      "witnessing": 35625,
      "pushing": 26895,
      "imperative": 14153,
      "inevitably": 14753,
      "detrimental": 7904,
      "environmental": 9666,
      "scant": 29186,
      "paid": 23967,
      "submodular": 31435,
      "biobert": 3738,
      "providers": 26745,
      "thousands": 33085,
      "coldstart": 4900,
      "puts": 26896,
      "tfew": 33042,
      "logic": 19154,
      "tl": 33181,
      "rigorously": 28886,
      "specify": 30767,
      "specification": 30762,
      "publish": 26869,
      "28k": 291,
      "lifted": 17825,
      "atomic": 2680,
      "ap": 2096,
      "originates": 23728,
      "usage": 34501,
      "characterizes": 4431,
      "logical": 19155,
      "richness": 28878,
      "ui": 34090,
      "pivotal": 24976,
      "digital": 8188,
      "facilitating": 10851,
      "encapsulating": 9335,
      "graphical": 13233,
      "guis": 13362,
      "gui": 13340,
      "mobile": 20318,
      "apps": 2423,
      "extensibility": 10669,
      "adaptability": 1170,
      "wikihow": 35601,
      "app": 2108,
      "agentlm": 1565,
      "mechanisms": 19753,
      "faithful": 10928,
      "uncovered": 34135,
      "alignments": 1785,
      "bruteforce": 3900,
      "boundless": 3845,
      "solves": 30504,
      "implementing": 14169,
      "boolean": 3812,
      "faithfully": 10929,
      "begun": 3308,
      "revolutionize": 28842,
      "sciences": 29246,
      "methodological": 19988,
      "personalize": 24887,
      "periods": 24860,
      "1972": 194,
      "alpaca7b": 1835,
      "missing": 20223,
      "entirely": 9628,
      "changed": 4408,
      "rising": 28891,
      "sociodemographic": 30441,
      "privacy": 25950,
      "broaden": 3888,
      "revolutionary": 28841,
      "drastically": 8721,
      "deficiency": 7361,
      "evident": 10070,
      "demanding": 7412,
      "sustained": 31787,
      "counseling": 6288,
      "synthesizing": 31843,
      "anthropomorphic": 2091,
      "incorporates": 14571,
      "updating": 34477,
      "permits": 24867,
      "forget": 11729,
      "closedsource": 4686,
      "exemplify": 10206,
      "llmbased": 18384,
      "chatbot": 4448,
      "acts": 1154,
      "recall": 27468,
      "querying": 27035,
      "600": 459,
      "043": 9,
      "tau": 32571,
      "compromises": 5438,
      "intents": 15378,
      "json": 15719,
      "dbpedia": 7205,
      "lima": 17906,
      "trip": 33911,
      "history": 13729,
      "tends": 32713,
      "strictly": 31146,
      "43": 388,
      "bard": 3095,
      "davinci003": 7200,
      "pseudocode": 26819,
      "harnessing": 13458,
      "132": 112,
      "aggregate": 1576,
      "comments": 4981,
      "docstrings": 8500,
      "assumptions": 2664,
      "breaks": 3860,
      "flan": 11586,
      "fairly": 10925,
      "adopted": 1404,
      "redundant": 27767,
      "nowadays": 23131,
      "simplifying": 30181,
      "wordbyword": 35653,
      "rerankers": 28277,
      "proposals": 26488,
      "t5small": 31978,
      "xsum": 35881,
      "mauve": 19697,
      "booksum": 3811,
      "paragraphlevel": 24163,
      "alms": 1819,
      "paraphrases": 24305,
      "traditionally": 33354,
      "falcon40b": 10945,
      "perception": 24461,
      "scheduling": 29229,
      "llmempowered": 18392,
      "unprecedented": 34418,
      "harnesses": 13456,
      "begins": 3307,
      "perceive": 24455,
      "microbatches": 20156,
      "llamabased": 18236,
      "86": 577,
      "compromising": 5439,
      "toolkits": 33266,
      "flashattention": 11609,
      "garnered": 12119,
      "handling": 13413,
      "reframe": 27803,
      "consumed": 5816,
      "reflected": 27793,
      "slu": 30334,
      "close": 4671,
      "oracle": 23663,
      "note": 23034,
      "polyglot": 25097,
      "encyclopedic": 9406,
      "associations": 2656,
      "counterfactuals": 6298,
      "metas": 19862,
      "difficulties": 8175,
      "location": 19148,
      "exploits": 10562,
      "monte": 22214,
      "carlo": 4183,
      "policies": 25078,
      "multiplication": 22433,
      "travel": 33868,
      "modelbased": 20879,
      "inadequate": 14437,
      "computes": 5510,
      "biographies": 3740,
      "commercial": 4982,
      "retrievalaugmented": 28760,
      "llmpowered": 18396,
      "utilise": 34949,
      "dollyv2": 8549,
      "stablevicuna": 30852,
      "synthesised": 31838,
      "translated": 33812,
      "coherence": 4886,
      "tamil": 32041,
      "falls": 10955,
      "personalized": 24888,
      "defects": 7358,
      "traversal": 33869,
      "enrich": 9591,
      "cover": 6314,
      "entries": 9660,
      "counteract": 6293,
      "enlarged": 9581,
      "add": 1227,
      "l1": 15949,
      "precomputed": 25396,
      "link": 18027,
      "coverage": 6318,
      "boosting": 3819,
      "expands": 10342,
      "globally": 12798,
      "western": 35535,
      "culture": 6461,
      "unfairness": 34305,
      "suited": 31602,
      "hindi": 13721,
      "112": 77,
      "nonenglish": 22990,
      "showcasing": 29847,
      "narratives": 22492,
      "claimed": 4578,
      "originating": 23729,
      "indices": 14708,
      "predicate": 25403,
      "conform": 5667,
      "controls": 6068,
      "hallucinate": 13367,
      "contextaware": 5930,
      "pay": 24424,
      "unfaithful": 34306,
      "amplifies": 1895,
      "difference": 8034,
      "143": 137,
      "overriding": 23951,
      "contradicts": 6006,
      "resolving": 28404,
      "conflict": 5664,
      "window": 35611,
      "segments": 29373,
      "plaintext": 24992,
      "precomputing": 25397,
      "passage": 24378,
      "inexpensive": 14754,
      "openworld": 23560,
      "profiles": 26188,
      "considers": 5727,
      "predefined": 25400,
      "ontology": 23377,
      "displaying": 8317,
      "emerges": 9186,
      "gpt35turbo": 13041,
      "bloomz": 3796,
      "whisper": 35538,
      "33": 329,
      "61": 462,
      "texttospeech": 33008,
      "tts": 33952,
      "resulted": 28551,
      "overarching": 23918,
      "exceptions": 10177,
      "gaps": 12116,
      "cornerstone": 6162,
      "motion": 22225,
      "subjectivity": 31426,
      "spearman": 30651,
      "nearly": 22602,
      "19k": 196,
      "grounds": 13292,
      "inhouse": 14957,
      "llama7b": 18230,
      "trustworthy": 33935,
      "entails": 9616,
      "checking": 4517,
      "pushed": 26894,
      "frontier": 11931,
      "ats": 2682,
      "exciting": 10183,
      "simplifications": 30176,
      "cheaply": 4511,
      "weaker": 35458,
      "selfinstruct": 29423,
      "canonical": 3998,
      "discrepancies": 8277,
      "mimicking": 20171,
      "closed": 4674,
      "shortcut": 29823,
      "paraphrasing": 24307,
      "distills": 8359,
      "extremescale": 10804,
      "hypothesize": 13966,
      "proximity": 26798,
      "occupy": 23272,
      "proximal": 26793,
      "subspace": 31456,
      "subspaces": 31457,
      "unconstrained": 34133,
      "fidelity": 11131,
      "assumed": 2660,
      "whitebox": 35541,
      "incompatible": 14536,
      "blackboxes": 3758,
      "opt30b": 23608,
      "observing": 23242,
      "23x": 261,
      "primed": 25925,
      "pronoun": 26455,
      "referential": 27774,
      "stimuli": 31064,
      "replicating": 28109,
      "johnson": 15709,
      "icl": 13973,
      "adapts": 1224,
      "flanul2": 11607,
      "contemporary": 5841,
      "verifiers": 35214,
      "violates": 35279,
      "actions": 1132,
      "verifier": 35213,
      "invalid": 15561,
      "check": 4513,
      "presence": 25507,
      "prune": 26802,
      "temperature": 32686,
      "curse": 6550,
      "descriptive": 7693,
      "astonishing": 2667,
      "ecosystem": 8814,
      "irreversible": 15647,
      "tails": 32021,
      "disappear": 8253,
      "collapse": 4909,
      "autoencoders": 2845,
      "intuition": 15557,
      "seriously": 29645,
      "sustain": 31784,
      "wealth": 35467,
      "referred": 27775,
      "selfknowledge": 29426,
      "paramount": 24298,
      "unanswerable": 34107,
      "selfaware": 29410,
      "discovering": 8272,
      "distinguishable": 8374,
      "sampleefficient": 29069,
      "templates": 32691,
      "aligning": 1744,
      "logit": 19161,
      "testtime": 32811,
      "nearest": 22600,
      "neighbors": 22675,
      "succeed": 31505,
      "linearly": 17999,
      "retrieving": 28780,
      "iteration": 15681,
      "establishes": 9775,
      "reflective": 27797,
      "categorizing": 4230,
      "typology": 34086,
      "conceptualize": 5536,
      "restricts": 28537,
      "macrof1": 19378,
      "pilot": 24954,
      "dolly": 8548,
      "openassistant": 23447,
      "stablelm": 30851,
      "deliberately": 7398,
      "inserted": 15057,
      "verifying": 35220,
      "respective": 28453,
      "sections": 29340,
      "neurips": 22764,
      "119": 83,
      "866": 579,
      "clearly": 4649,
      "distinctions": 8371,
      "think": 33065,
      "meaningfully": 19726,
      "vocabularies": 35380,
      "incomparable": 14535,
      "playing": 25024,
      "modelagnostic": 20878,
      "similarities": 30123,
      "differences": 8036,
      "believed": 3343,
      "customize": 6563,
      "backward": 3070,
      "stores": 31083,
      "tensorized": 32724,
      "orca": 23665,
      "progressive": 26232,
      "traces": 33324,
      "thought": 33078,
      "tap": 32044,
      "judicious": 15729,
      "vicuna13b": 35258,
      "42": 386,
      "agieval": 1584,
      "pts": 26828,
      "examinations": 10097,
      "sat": 29103,
      "lsat": 19324,
      "gre": 13246,
      "indicates": 14699,
      "modelsllms": 22146,
      "followers": 11684,
      "forbidden": 11715,
      "lowdimensional": 19280,
      "converted": 6113,
      "children": 4533,
      "categorize": 4228,
      "lexicon": 17805,
      "master": 19636,
      "principles": 25933,
      "aged": 1556,
      "months": 22216,
      "96": 605,
      "triggering": 33900,
      "memorizing": 19799,
      "chain": 4289,
      "upper": 34481,
      "byproduct": 3952,
      "sql": 30838,
      "lambda": 16020,
      "calculus": 3971,
      "impeding": 14152,
      "164": 161,
      "alongside": 1820,
      "generaldomain": 12190,
      "55b": 447,
      "protocol": 26662,
      "gather": 12126,
      "equip": 9678,
      "naive": 22477,
      "llmasajudge": 18382,
      "mtbench": 22259,
      "arena": 2487,
      "inadequacy": 14436,
      "judges": 15724,
      "verbosity": 35206,
      "agreement": 1590,
      "battle": 3293,
      "complement": 5243,
      "30k": 314,
      "specialists": 30661,
      "englishcentric": 9495,
      "transferability": 33685,
      "welltrained": 35531,
      "bea": 3298,
      "teacherstudent": 32589,
      "participated": 24329,
      "codalab": 4719,
      "experimented": 10413,
      "opt27b": 23607,
      "bertscore": 3550,
      "dialogrpt": 8007,
      "pedagogical": 24431,
      "achievements": 1023,
      "participating": 24330,
      "ais": 1674,
      "googles": 12831,
      "mediate": 19764,
      "moderation": 22154,
      "worlds": 35836,
      "7000": 499,
      "attempted": 2706,
      "accounts": 853,
      "policymakers": 25091,
      "telecom": 32682,
      "frontiers": 11934,
      "automating": 2924,
      "evolution": 10074,
      "realizing": 27323,
      "3rd": 370,
      "partnership": 24364,
      "pertinent": 24906,
      "846": 574,
      "83": 568,
      "corroborates": 6235,
      "stepping": 31056,
      "wireless": 35620,
      "paves": 24418,
      "revolution": 28840,
      "private": 25957,
      "harmful": 13439,
      "unauthorized": 34110,
      "inclusion": 14532,
      "copyrighted": 6147,
      "repositories": 28129,
      "gpts": 13163,
      "truly": 33928,
      "apache": 2097,
      "licenses": 17816,
      "entry": 9663,
      "hurdles": 13950,
      "openness": 23475,
      "thoughts": 33082,
      "formalized": 11749,
      "openai": 23432,
      "leaking": 17500,
      "distant": 8328,
      "compiling": 5242,
      "problemsolving": 26037,
      "formalizing": 11750,
      "received": 27476,
      "lowlevel": 19297,
      "bridging": 3873,
      "prowess": 26790,
      "vital": 35370,
      "languagespecific": 16923,
      "instructing": 15138,
      "89": 582,
      "multitude": 22459,
      "homepage": 13750,
      "section": 29339,
      "belongs": 3347,
      "titles": 33179,
      "layout": 17454,
      "paraphraser": 24304,
      "substitutions": 31494,
      "arxiv": 2547,
      "peer": 24432,
      "reviewed": 28831,
      "mse": 22249,
      "scibert": 29239,
      "alterations": 1843,
      "regularizes": 27829,
      "gradual": 13200,
      "rank": 27221,
      "arbitrary": 2428,
      "ranks": 27236,
      "fullrank": 11950,
      "36": 353,
      "analyzes": 2000,
      "polarity": 25077,
      "reversed": 28824,
      "integrated": 15323,
      "calculated": 3962,
      "textbooks": 32977,
      "phi1": 24927,
      "competing": 5215,
      "a100s": 620,
      "textbook": 32976,
      "6b": 489,
      "exercises": 10208,
      "1b": 197,
      "attains": 2698,
      "pass1": 24372,
      "humaneval": 13896,
      "mbpp": 19714,
      "coding": 4864,
      "350m": 343,
      "savings": 29117,
      "illustrate": 14045,
      "accelerates": 796,
      "empowering": 9268,
      "untapped": 34459,
      "uncovers": 34138,
      "expressing": 10641,
      "overconfidence": 23929,
      "narrow": 22493,
      "auroc": 2831,
      "scope": 29263,
      "composing": 5323,
      "prevailing": 25854,
      "integrating": 15330,
      "negation": 22657,
      "operators": 23576,
      "unlearning": 34387,
      "alpacalora": 1839,
      "authoring": 2835,
      "practically": 25376,
      "oriented": 23696,
      "opportunities": 23580,
      "cheating": 4512,
      "measurable": 19731,
      "consuming": 5822,
      "127": 100,
      "practitioner": 25380,
      "sophistication": 30528,
      "leveraged": 17764,
      "designers": 7750,
      "cfg": 4287,
      "texttoimage": 33005,
      "pure": 26879,
      "pythia": 26899,
      "llamafamily": 18240,
      "twice": 34036,
      "stack": 30854,
      "selfconsistency": 29413,
      "75": 515,
      "gpt4all": 13128,
      "tensortrain": 32726,
      "highdimensional": 13591,
      "subtle": 31498,
      "dimensionality": 8197,
      "lies": 17819,
      "susceptibility": 31780,
      "trust": 33932,
      "quantification": 26984,
      "hurdle": 13949,
      "roadblock": 28913,
      "representativeness": 28188,
      "redundancy": 27766,
      "suffice": 31558,
      "lengthy": 17716,
      "regrettably": 27824,
      "methodologies": 19989,
      "equal": 9673,
      "disregarding": 8319,
      "inequalities": 14751,
      "receive": 27475,
      "excessive": 10178,
      "weighting": 35500,
      "rectify": 27674,
      "wizardlm": 35626,
      "llama2chat": 18215,
      "33b": 332,
      "freeform": 11912,
      "came": 3986,
      "overheads": 23939,
      "memoryefficient": 19837,
      "penalty": 24448,
      "goals": 12810,
      "32768": 323,
      "hinder": 13714,
      "accessibility": 831,
      "ls": 19323,
      "sharedtask": 29787,
      "portuguese": 25152,
      "nextword": 22898,
      "excelled": 10159,
      "suboptimal": 31438,
      "delivering": 7401,
      "answerability": 2053,
      "355m": 348,
      "em": 9118,
      "geometry": 12728,
      "spatial": 30637,
      "llmsgenerated": 19052,
      "geometric": 12725,
      "spatially": 30644,
      "highlights": 13649,
      "complexities": 5299,
      "organized": 23693,
      "simulating": 30189,
      "measured": 19738,
      "contributing": 6034,
      "matched": 19648,
      "misleadingly": 20219,
      "recognized": 27645,
      "superposition": 31668,
      "contextdependent": 5931,
      "opposed": 23584,
      "affordance": 1550,
      "differing": 8166,
      "implied": 14181,
      "smoothness": 30412,
      "outline": 23742,
      "website": 35487,
      "developments": 7978,
      "options": 23662,
      "positional": 25185,
      "moderatesized": 22153,
      "debiasing": 7215,
      "exploration": 10563,
      "decide": 7227,
      "gsm8k": 13325,
      "suffered": 31555,
      "llmdriven": 18391,
      "agent": 1558,
      "ahead": 1592,
      "decomposing": 7288,
      "snippets": 30414,
      "htmlt5": 13773,
      "mind2web": 20175,
      "graphtotext": 13240,
      "agenda": 1557,
      "webnlg": 35482,
      "transportation": 33865,
      "chatgptgenerated": 4504,
      "instructionoutput": 15237,
      "sft": 29757,
      "rlhf": 28906,
      "hindering": 13716,
      "instructiontune": 15282,
      "left": 17691,
      "raised": 27164,
      "26": 278,
      "learningbased": 17673,
      "aimed": 1650,
      "harms": 13451,
      "moral": 22218,
      "expertlevel": 10517,
      "gemini": 12138,
      "pro": 25961,
      "anthropics": 2089,
      "claude": 4635,
      "70b": 503,
      "shifts": 29807,
      "tendency": 32711,
      "recommends": 27657,
      "caution": 4257,
      "distinctive": 8372,
      "slightly": 30322,
      "bot": 3835,
      "sparkdesk": 30603,
      "chatgpt35": 4500,
      "aids": 1630,
      "convincing": 6119,
      "calculations": 3967,
      "chart": 4437,
      "vqa": 35394,
      "nonsynthetic": 23011,
      "charts": 4438,
      "palm2": 23993,
      "openvocabulary": 23558,
      "textonly": 32984,
      "mentioning": 19844,
      "rating": 27272,
      "0shot": 23,
      "mllm": 20292,
      "llava": 18242,
      "mplugowl": 22243,
      "blip2": 3775,
      "openflamingos": 23465,
      "llava13b": 18250,
      "cider": 4565,
      "serialized": 29630,
      "llavas": 18256,
      "015": 3,
      "reaching": 27288,
      "026": 7,
      "coupled": 6309,
      "speculative": 30777,
      "ondevice": 23338,
      "intensity": 15369,
      "rationale": 27276,
      "explanatory": 10543,
      "utilising": 34950,
      "strategyqa": 31131,
      "617": 463,
      "636": 471,
      "727": 510,
      "316": 316,
      "255": 273,
      "extrapolating": 10788,
      "laws": 17420,
      "allocation": 1799,
      "xquad": 35880,
      "flores101": 11626,
      "encouragingly": 9405,
      "middle": 20158,
      "defined": 7364,
      "benefiting": 3488,
      "standalone": 30866,
      "categorical": 4220,
      "crossentropy": 6409,
      "oneatatime": 23341,
      "leaving": 17681,
      "synonyms": 31812,
      "streamlines": 31136,
      "spatiotemporal": 30645,
      "coordinates": 6137,
      "encapsulate": 9332,
      "interval": 15474,
      "deception": 7225,
      "claim": 4576,
      "editing": 8825,
      "ultimately": 34097,
      "deceptive": 7226,
      "risen": 28889,
      "precise": 25388,
      "088": 19,
      "85": 575,
      "pegasus": 24444,
      "82": 565,
      "57": 449,
      "liar": 17810,
      "alter": 1842,
      "preserves": 25604,
      "probed": 25977,
      "embarks": 9124,
      "pioneering": 24959,
      "commonlyused": 5032,
      "detectors": 7888,
      "scrutinizing": 29297,
      "unveil": 34466,
      "detector": 7887,
      "intriguing": 15487,
      "isotropic": 15651,
      "anisotropic": 2012,
      "succinct": 31547,
      "gptbased": 13139,
      "humanauthored": 13887,
      "closing": 4700,
      "foreign": 11724,
      "purposes": 26885,
      "ensuring": 9609,
      "inefficiency": 14749,
      "ineffective": 14747,
      "affordable": 1549,
      "transformerlike": 33773,
      "openllama": 23470,
      "3b": 364,
      "ensembling": 9599,
      "fundamentally": 11983,
      "foundations": 11810,
      "realized": 27321,
      "reality": 27319,
      "unravel": 34422,
      "cooperation": 6133,
      "knowledgeguided": 15928,
      "discovered": 8269,
      "variation": 35044,
      "concatenation": 5523,
      "mainstream": 19408,
      "069": 13,
      "048": 10,
      "comet": 4976,
      "enhancements": 9543,
      "blue": 3798,
      "leakage": 17498,
      "lacks": 16012,
      "subjective": 31424,
      "dimensions": 8199,
      "advent": 1508,
      "kb": 15741,
      "llama213bchat": 18200,
      "005": 1,
      "fetched": 11082,
      "dpr": 8709,
      "held": 13497,
      "recommendation": 27648,
      "60k": 461,
      "glove": 12799,
      "passed": 24382,
      "collaborate": 4902,
      "fedllm": 11052,
      "federated": 11048,
      "fl": 11583,
      "addressed": 1360,
      "integration": 15339,
      "bpfree": 3849,
      "perturbed": 24912,
      "delivers": 7402,
      "centers": 4265,
      "designs": 7755,
      "striking": 31152,
      "careful": 4167,
      "balance": 3077,
      "uniquely": 34365,
      "cots": 6287,
      "unattained": 34109,
      "necessitating": 22613,
      "compilers": 5241,
      "compiler": 5240,
      "passes": 24383,
      "passing": 24384,
      "133": 113,
      "count": 6289,
      "104": 58,
      "food": 11709,
      "autonomously": 2933,
      "blip": 3774,
      "customization": 6562,
      "underscoring": 34184,
      "bringing": 3878,
      "centered": 4264,
      "encompass": 9385,
      "phonetics": 24935,
      "631": 469,
      "llama270bchat": 18203,
      "android": 2010,
      "smartphones": 30409,
      "developers": 7937,
      "endusers": 9445,
      "bridges": 3871,
      "multigranularity": 22283,
      "158": 149,
      "909": 594,
      "713": 508,
      "gpt4powered": 13131,
      "364": 354,
      "397": 362,
      "suites": 31603,
      "marked": 19592,
      "preprocessing": 25505,
      "analytics": 1985,
      "delve": 7406,
      "imputation": 14428,
      "contextualization": 5959,
      "immense": 14113,
      "underscores": 34179,
      "dnns": 8498,
      "depth": 7666,
      "width": 35600,
      "submodels": 31434,
      "nested": 22682,
      "submodel": 31433,
      "accumulation": 855,
      "versatility": 35229,
      "validated": 34998,
      "mobilenet": 20320,
      "superiority": 31661,
      "awareness": 3044,
      "unexpectedly": 34298,
      "llama1": 18153,
      "pytorch": 26905,
      "prioritizing": 25948,
      "communitys": 5063,
      "implementations": 14164,
      "codebase": 4831,
      "highperforming": 13680,
      "confined": 5658,
      "xgen": 35868,
      "8k": 584,
      "instructional": 15211,
      "dollars": 8547,
      "consolidate": 5767,
      "glm130b": 12792,
      "contrasting": 6019,
      "deviates": 7986,
      "later": 17413,
      "localized": 19140,
      "justify": 15736,
      "definitive": 7370,
      "incorrectly": 14590,
      "hampers": 13401,
      "trustworthiness": 33934,
      "answered": 2054,
      "vietnamese": 35271,
      "consideration": 5714,
      "helpfulness": 13518,
      "2030": 241,
      "fake": 10933,
      "peftlora": 24439,
      "analysing": 1908,
      "manipulation": 19543,
      "neurons": 22768,
      "dead": 7207,
      "125m": 97,
      "66b": 481,
      "neuron": 22765,
      "70": 494,
      "activate": 1133,
      "stream": 31132,
      "acting": 1129,
      "phi15": 24928,
      "initiated": 14974,
      "5x": 457,
      "seeing": 29356,
      "vertical": 35239,
      "interested": 15405,
      "subdomains": 31420,
      "deeply": 7351,
      "driving": 8737,
      "tendencies": 32710,
      "favoring": 11013,
      "lives": 18056,
      "disparities": 8312,
      "lean": 17502,
      "elucidate": 9114,
      "circuit": 4567,
      "breaking": 3859,
      "undesirable": 34290,
      "pathways": 24404,
      "intention": 15374,
      "disabling": 8247,
      "behaves": 3310,
      "ablate": 732,
      "edges": 8821,
      "characterizing": 4433,
      "figures": 11161,
      "characterization": 4428,
      "argument": 2492,
      "astronomy": 2671,
      "scholarly": 29236,
      "7billionparameter": 545,
      "insightful": 15067,
      "stateofthearts": 31009,
      "spur": 30831,
      "equivalence": 9686,
      "ontologies": 23376,
      "flant5xxl": 11606,
      "selfassessment": 29406,
      "tried": 33896,
      "measurement": 19739,
      "administering": 1397,
      "template": 32689,
      "draft": 8710,
      "twostage": 34043,
      "drafting": 8713,
      "skipping": 30318,
      "ensures": 9608,
      "identical": 13991,
      "opened": 23453,
      "trial": 33892,
      "connecting": 5683,
      "borrows": 3834,
      "starts": 30901,
      "population": 25147,
      "31": 315,
      "synergies": 31807,
      "inspire": 15090,
      "costefficient": 6260,
      "expanding": 10339,
      "really": 27326,
      "gptneox": 13159,
      "20b": 246,
      "formatting": 11760,
      "adherence": 1385,
      "excelling": 10163,
      "integral": 15318,
      "minimizing": 20199,
      "triviaqa": 33921,
      "earlyexit": 8783,
      "trainingfree": 33649,
      "li": 17807,
      "searches": 29314,
      "2l": 298,
      "hellaswag": 13499,
      "highlyefficient": 13671,
      "highperformance": 13674,
      "restrictive": 28536,
      "multiplications": 22434,
      "underutilized": 34288,
      "onchip": 23337,
      "kernel": 15749,
      "15x": 155,
      "36x": 357,
      "fastertransformer": 11004,
      "baichuan": 3075,
      "cmmlu": 4710,
      "excels": 10164,
      "observational": 23220,
      "activates": 1136,
      "denoted": 7603,
      "gpt4generated": 13130,
      "confident": 5652,
      "rates": 27270,
      "tied": 33104,
      "argued": 2491,
      "imply": 14183,
      "matters": 19696,
      "implies": 14182,
      "care": 4165,
      "economic": 8810,
      "unpredictable": 34421,
      "impactful": 14144,
      "credible": 6374,
      "promotes": 26302,
      "craft": 6339,
      "scaffold": 29121,
      "feeding": 11078,
      "llama213b": 18197,
      "path": 24397,
      "slimpajama": 30324,
      "627b": 465,
      "weve": 35536,
      "deduplication": 7314,
      "proportions": 26486,
      "alibi": 1723,
      "swiglu": 31793,
      "configuration": 5654,
      "bf16": 3638,
      "specializing": 30678,
      "700": 498,
      "rephrased": 28097,
      "drawbacks": 8725,
      "t53b": 31968,
      "stateofart": 30916,
      "exposed": 10632,
      "senses": 29510,
      "nllb": 22914,
      "localizing": 19142,
      "devoted": 7996,
      "imbued": 14104,
      "inadequately": 14438,
      "rlaif": 28904,
      "reward": 28857,
      "attuned": 2770,
      "valuealigned": 35025,
      "applicationspecific": 2184,
      "alpacaeval": 1837,
      "turbo": 34023,
      "quantized": 27012,
      "brazilian": 3856,
      "school": 29238,
      "home": 13749,
      "database": 6918,
      "took": 33255,
      "amd": 1876,
      "processor": 26136,
      "stateful": 30913,
      "mock": 20321,
      "roles": 28964,
      "triggers": 33901,
      "monologue": 22212,
      "assesses": 2607,
      "necessitate": 22611,
      "calculation": 3965,
      "deemed": 7315,
      "calculationintensive": 3966,
      "reversal": 28819,
      "expose": 10631,
      "deduction": 7308,
      "composer": 5321,
      "melodies": 19788,
      "lee": 17690,
      "79": 523,
      "card": 4162,
      "rise": 28887,
      "documenting": 8514,
      "pitfalls": 24974,
      "bengali": 3493,
      "modest": 22177,
      "regard": 27806,
      "claude2": 4638,
      "calls": 3984,
      "tablebased": 31985,
      "flant5base": 11602,
      "defining": 7366,
      "psychometric": 26824,
      "dilemmas": 8194,
      "considerations": 5715,
      "pointing": 25068,
      "splitting": 30812,
      "differentiable": 8161,
      "kernels": 15751,
      "floating": 11621,
      "upto": 34488,
      "starcoder": 30894,
      "16b": 164,
      "minutes": 20206,
      "nearlossless": 22601,
      "spqr": 30826,
      "qlora": 26922,
      "32": 318,
      "manipulate": 19539,
      "persons": 24894,
      "equals": 9675,
      "unless": 34391,
      "irrespective": 15645,
      "extractable": 10747,
      "adequate": 1381,
      "instruct": 15128,
      "promptlearning": 26402,
      "prominence": 26263,
      "slms": 30327,
      "customizability": 6560,
      "costeffectiveness": 6259,
      "retail": 28715,
      "slm": 30326,
      "underscore": 34172,
      "exceeding": 10149,
      "leap": 17503,
      "emphasizing": 9215,
      "dearth": 7210,
      "collegelevel": 4942,
      "universitys": 34377,
      "gpt354": 13040,
      "shortcomings": 29821,
      "affecting": 1541,
      "pace": 23961,
      "underinvestigated": 34150,
      "parse": 24315,
      "embodying": 9151,
      "commercially": 4999,
      "workflows": 35802,
      "apparent": 2109,
      "weakness": 35463,
      "adhering": 1386,
      "textrelated": 32986,
      "encounter": 9391,
      "argumentative": 2493,
      "essays": 9753,
      "boasting": 3802,
      "cohen": 4884,
      "kappa": 15739,
      "spectrum": 30771,
      "meticulously": 20113,
      "ratings": 27273,
      "evaluator": 10042,
      "referencefree": 27772,
      "universal": 34370,
      "existed": 10258,
      "standardized": 30886,
      "duration": 8753,
      "trail": 33358,
      "timesensitive": 33167,
      "predominant": 25460,
      "engage": 9452,
      "forecasting": 11722,
      "occurrence": 23275,
      "actionable": 1131,
      "bag": 3074,
      "transcribed": 33664,
      "aggregated": 1577,
      "visualized": 35363,
      "academia": 782,
      "worth": 35842,
      "underperform": 34165,
      "unveiled": 34467,
      "llama27b": 18204,
      "intricate": 15483,
      "portrayal": 25151,
      "showcased": 29843,
      "clinical": 4656,
      "consultation": 5814,
      "distinction": 8370,
      "manifestation": 19536,
      "personalities": 24883,
      "thirteen": 33069,
      "classifies": 4629,
      "textdavinci003": 32980,
      "jailbreak": 15695,
      "bypass": 3947,
      "protocols": 26663,
      "surrogate": 31769,
      "llmss": 19053,
      "deduce": 7306,
      "generalizing": 12237,
      "intelligently": 15365,
      "circumstances": 4570,
      "september": 29586,
      "78": 522,
      "responds": 28470,
      "inconsistency": 14539,
      "validating": 35003,
      "delves": 7407,
      "multiagent": 22265,
      "embodied": 9150,
      "sparked": 30604,
      "debate": 7211,
      "superficial": 31640,
      "2d": 296,
      "bertlike": 3544,
      "nuggets": 23137,
      "exploded": 10552,
      "meta": 19856,
      "multinode": 22372,
      "multigpu": 22281,
      "sharding": 29779,
      "subnetworks": 31437,
      "disentangling": 8307,
      "multiobjective": 22373,
      "adverse": 1523,
      "uncovering": 34136,
      "maintains": 19431,
      "socratic": 30444,
      "hampered": 13400,
      "buggy": 3905,
      "instructionbased": 15214,
      "planners": 24998,
      "formulating": 11775,
      "stacking": 30856,
      "hardcoded": 13424,
      "invoked": 15622,
      "collecting": 4924,
      "compositions": 5330,
      "loops": 19223,
      "546": 445,
      "chains": 4306,
      "decade": 7218,
      "surge": 31724,
      "genai": 12147,
      "prospects": 26654,
      "earlystage": 8784,
      "foster": 11785,
      "whats": 35537,
      "magic": 19380,
      "formalize": 11748,
      "dynamical": 8764,
      "steers": 31031,
      "singular": 30236,
      "complementary": 5244,
      "falcon7b": 10946,
      "wikitext": 35605,
      "97": 606,
      "intriguingly": 15488,
      "suppression": 31721,
      "comprehensively": 5396,
      "107": 61,
      "explains": 10534,
      "favored": 11012,
      "selfrepair": 29433,
      "suppress": 31720,
      "39": 361,
      "indonesia": 14728,
      "manages": 19531,
      "distantlysupervised": 8329,
      "arises": 2497,
      "impair": 14148,
      "penalizes": 24446,
      "entityrelation": 9657,
      "coarsegrained": 4715,
      "dictated": 8027,
      "disciplines": 8259,
      "impede": 14150,
      "technological": 32675,
      "terminologies": 32734,
      "aiassisted": 1624,
      "spotlight": 30821,
      "v1": 34980,
      "categorized": 4229,
      "hierarchies": 13546,
      "equity": 9685,
      "assignments": 2628,
      "underrepresented": 34169,
      "proposal": 26487,
      "neftune": 22653,
      "rises": 28890,
      "evolinstruct": 10073,
      "sharegpt": 29788,
      "satellite": 29104,
      "embed": 9125,
      "geographic": 12724,
      "density": 7612,
      "pearsons": 24430,
      "crucially": 6452,
      "agriculture": 1591,
      "healthcare": 13487,
      "certification": 4283,
      "rag": 27154,
      "india": 14681,
      "gpt4s": 13132,
      "earn": 8785,
      "certifications": 4284,
      "admission": 1399,
      "management": 19529,
      "professionals": 26177,
      "trace": 33323,
      "continual": 5973,
      "tasksolving": 32558,
      "effortless": 9081,
      "declines": 7244,
      "288": 289,
      "losses": 19253,
      "recovery": 27671,
      "executed": 10194,
      "cpu": 6335,
      "runtimes": 29018,
      "memorybound": 19836,
      "recovering": 27669,
      "validates": 35002,
      "replication": 28110,
      "customizing": 6566,
      "highvolume": 13713,
      "untargeted": 34460,
      "selfrefinement": 29429,
      "devoid": 7995,
      "strikingly": 31153,
      "footprints": 11714,
      "vicuna7b": 35260,
      "ended": 9425,
      "domainagnostic": 8608,
      "string": 31154,
      "digits": 8192,
      "frame": 11820,
      "extrapolate": 10787,
      "purposebuilt": 26884,
      "tokenizing": 33213,
      "repetition": 28093,
      "daytoday": 7203,
      "personalization": 24886,
      "surpassed": 31736,
      "backed": 3060,
      "specialize": 30663,
      "proactively": 25965,
      "competence": 5211,
      "lacking": 16011,
      "rubric": 29004,
      "20k": 247,
      "evaluators": 10043,
      "bench": 3348,
      "bolstering": 3808,
      "hhh": 13534,
      "honeybee": 13752,
      "loop": 19222,
      "progressively": 26233,
      "outperformance": 23790,
      "queried": 27017,
      "gametheoretic": 12084,
      "casts": 4210,
      "seeks": 29360,
      "progression": 26231,
      "detailing": 7843,
      "rigor": 28882,
      "306": 312,
      "literary": 18039,
      "correlates": 6215,
      "arbitrarily": 2427,
      "binding": 3736,
      "mcqa": 19716,
      "formulas": 11769,
      "chemistry": 4527,
      "biology": 3743,
      "typed": 34056,
      "llama270b": 18201,
      "deriving": 7675,
      "firstofitskind": 11565,
      "24k": 267,
      "wave": 35427,
      "enterprise": 9619,
      "societal": 30436,
      "mirroring": 20208,
      "opacity": 23379,
      "reversing": 28826,
      "upstream": 34486,
      "standardize": 30885,
      "developer": 7936,
      "market": 19598,
      "standards": 30888,
      "regulatory": 27833,
      "h2o": 13363,
      "studio": 31292,
      "inherit": 14954,
      "rankings": 27235,
      "url": 34498,
      "httpsgithubcommicrosoftlmops": 13774,
      "selfimprove": 29421,
      "revise": 28834,
      "absent": 759,
      "selfimprovement": 29422,
      "replay": 28107,
      "reordering": 28085,
      "wordorder": 35654,
      "tags": 32012,
      "clause": 4641,
      "clauses": 4642,
      "intact": 15315,
      "embedded": 9126,
      "confirming": 5662,
      "implying": 14184,
      "discerning": 8257,
      "traintest": 33655,
      "wang": 35414,
      "draw": 8723,
      "categorization": 4227,
      "higherquality": 13616,
      "margins": 19590,
      "exception": 10165,
      "square": 30843,
      "beat": 3302,
      "adequately": 1382,
      "169": 163,
      "establishing": 9778,
      "mistral": 20227,
      "neglect": 22666,
      "advocate": 1530,
      "intricacies": 15482,
      "strengthening": 31140,
      "association": 2655,
      "axis": 3048,
      "connects": 5689,
      "merits": 19852,
      "anchor": 2004,
      "1100": 75,
      "newlyconstructed": 22873,
      "tuples": 34022,
      "snippet": 30413,
      "elaborating": 9094,
      "multifaceted": 22280,
      "satisfying": 29111,
      "inability": 14429,
      "branch": 3853,
      "merge": 19848,
      "humanllm": 13913,
      "satisfaction": 29105,
      "compresses": 5408,
      "db": 7204,
      "women": 35632,
      "pregnancy": 25486,
      "03": 8,
      "disagree": 8250,
      "causing": 4256,
      "mechanistic": 19755,
      "determinants": 7894,
      "capital": 4129,
      "london": 19165,
      "country": 6306,
      "city": 4574,
      "localize": 19139,
      "proof": 26459,
      "ada": 1159,
      "diluting": 8195,
      "balances": 3085,
      "occupational": 23270,
      "inclusive": 14533,
      "relates": 27861,
      "occupations": 23271,
      "request": 28206,
      "hierarchically": 13545,
      "occupation": 23269,
      "responsibility": 28518,
      "inquiries": 15054,
      "balanced": 3082,
      "tulu": 33953,
      "win": 35609,
      "864": 578,
      "groundbreaking": 13286,
      "wellestablished": 35523,
      "yielded": 35914,
      "seamlessly": 29303,
      "beats": 3303,
      "annealing": 2013,
      "trade": 33331,
      "modality": 20327,
      "amidst": 1881,
      "persists": 24877,
      "infeasible": 14755,
      "computationefficient": 5490,
      "scorer": 29276,
      "bind": 3735,
      "green": 13282,
      "shapes": 29777,
      "colors": 4944,
      "attaching": 2684,
      "distances": 8327,
      "expandable": 10337,
      "develops": 7984,
      "sizeable": 30290,
      "captured": 4152,
      "suggestive": 31589,
      "going": 12812,
      "sketch": 30305,
      "seeds": 29355,
      "swap": 31789,
      "pandalm": 23999,
      "5k": 455,
      "broadening": 3889,
      "selfcorrection": 29416,
      "inaccurately": 14435,
      "emulate": 9274,
      "aligns": 1786,
      "openllm": 23471,
      "demographic": 7428,
      "infrastructure": 14937,
      "astronomical": 2670,
      "identifiers": 14000,
      "maintenance": 19434,
      "identifier": 13999,
      "knowing": 15807,
      "nonfactual": 23000,
      "impedes": 14151,
      "diversify": 8475,
      "referring": 27776,
      "infinite": 14835,
      "approximations": 2422,
      "logistic": 19159,
      "marginally": 19589,
      "dnn": 8496,
      "burden": 3941,
      "likewise": 17905,
      "brought": 3894,
      "pushdown": 26893,
      "longtail": 19215,
      "tracks": 33329,
      "depths": 7667,
      "autoregressively": 2958,
      "softly": 30450,
      "modulate": 22189,
      "skip": 30317,
      "constituents": 5774,
      "silver": 30093,
      "35x": 352,
      "perplexities": 24871,
      "gpt2medium": 12975,
      "advantageous": 1502,
      "industrial": 14740,
      "parsers": 24317,
      "fulltraining": 11951,
      "backgrounds": 3063,
      "dissatisfaction": 8323,
      "posts": 25230,
      "divergent": 8407,
      "everincreasing": 10052,
      "divergence": 8406,
      "ip": 15639,
      "int8": 15314,
      "outlier": 23740,
      "deteriorated": 7890,
      "readers": 27299,
      "userspecified": 34705,
      "motivates": 22230,
      "emphasis": 9209,
      "reweighting": 28864,
      "directing": 8220,
      "meant": 19730,
      "posing": 25177,
      "coderelated": 4846,
      "codellama": 4841,
      "qwen": 27138,
      "744": 514,
      "67": 482,
      "customer": 6557,
      "unexpected": 34297,
      "correspondingly": 6232,
      "responding": 28469,
      "dpo": 8707,
      "introductory": 15556,
      "10000": 46,
      "customizable": 6561,
      "lunch": 19330,
      "assimilating": 2631,
      "delta": 7404,
      "disparity": 8313,
      "zeros": 35948,
      "fusing": 12020,
      "ranges": 27215,
      "effortlessly": 9082,
      "eliminate": 9106,
      "amalgamation": 1866,
      "wizardmath": 35628,
      "merged": 19849,
      "datacentric": 6922,
      "records": 27665,
      "pursue": 26886,
      "enlarging": 9582,
      "organization": 23690,
      "welldefined": 35520,
      "contamination": 5840,
      "overfit": 23931,
      "818": 563,
      "unintentional": 34352,
      "actively": 1150,
      "fresh": 11929,
      "labelled": 15963,
      "compounds": 5333,
      "organisms": 23689,
      "resourceintensive": 28426,
      "framing": 11908,
      "biogptlarge": 3739,
      "emulated": 9275,
      "predicts": 25458,
      "unexplored": 34299,
      "pubmed": 26873,
      "adhere": 1384,
      "authenticity": 2833,
      "emphasize": 9210,
      "refining": 27785,
      "optiml": 23659,
      "unfeasible": 34309,
      "severely": 29754,
      "promptsource": 26449,
      "cooperate": 6132,
      "adaptations": 1196,
      "causally": 4251,
      "negated": 22654,
      "selfpaced": 29427,
      "modalities": 20322,
      "81": 560,
      "llavav15": 18257,
      "vice": 35243,
      "versa": 35222,
      "facilitated": 10848,
      "encounters": 9396,
      "deterioration": 7893,
      "nles": 22908,
      "zephyr": 35931,
      "heuristics": 13533,
      "identically": 13992,
      "ought": 23733,
      "worst": 35840,
      "timeline": 33153,
      "reflecting": 27794,
      "unlabelled": 34385,
      "mixtures": 20288,
      "decode": 7245,
      "sequentiality": 29628,
      "altogether": 1859,
      "knearest": 15804,
      "knn": 15805,
      "elusive": 9116,
      "117m": 82,
      "functionalities": 11966,
      "educationally": 8847,
      "originate": 23727,
      "promptengineered": 26370,
      "qag": 26920,
      "ics": 13980,
      "flant5xl": 11605,
      "mistral7b": 20237,
      "sheds": 29799,
      "diagnosing": 7999,
      "unfiltered": 34310,
      "inadvertently": 14439,
      "reproduced": 28201,
      "interchangeably": 15398,
      "humanai": 13882,
      "receives": 27484,
      "prepare": 25498,
      "variablelength": 35036,
      "microbatch": 20155,
      "325x": 321,
      "proportionally": 26485,
      "aggressively": 1582,
      "regular": 27825,
      "invoke": 15621,
      "comprised": 5430,
      "alpaca52k": 1834,
      "double": 8667,
      "tokenlevel": 33214,
      "smallersized": 30403,
      "mixtureofexpert": 20283,
      "gradientfree": 13196,
      "repository": 28130,
      "bit": 3749,
      "unigram": 34348,
      "twostep": 34048,
      "competitiveness": 5234,
      "attaining": 2697,
      "contend": 5847,
      "thesis": 33063,
      "catered": 4234,
      "hpc": 13770,
      "concentrate": 5524,
      "managing": 19532,
      "race": 27141,
      "pave": 24416,
      "resilient": 28395,
      "merging": 19851,
      "bge": 3640,
      "mteb": 22261,
      "inversion": 15569,
      "preceding": 25385,
      "reconstructs": 27661,
      "recovers": 27670,
      "innovatively": 15001,
      "selfchat": 29411,
      "undergo": 34145,
      "resultant": 28549,
      "mastery": 19638,
      "disruptive": 8320,
      "integrity": 15346,
      "moss": 22224,
      "gptzero": 13166,
      "characterbased": 4423,
      "peoples": 24452,
      "desires": 7766,
      "identities": 14025,
      "viewpoints": 35276,
      "langauge": 16029,
      "40b": 379,
      "180b": 185,
      "assembled": 2584,
      "falcon180b": 10944,
      "documented": 8512,
      "dive": 8404,
      "tooling": 33264,
      "4096": 378,
      "cognition": 4872,
      "grasping": 13245,
      "aspire": 2581,
      "aviation": 3034,
      "abundance": 779,
      "jargon": 15701,
      "terminology": 32735,
      "opportunity": 23583,
      "nas": 22495,
      "separated": 29583,
      "door": 8666,
      "quest": 27036,
      "vehicles": 35199,
      "responsiveness": 28525,
      "visionlanguage": 35312,
      "openflamingo": 23464,
      "consolidated": 5768,
      "holistic": 13743,
      "instant": 15114,
      "reflection": 27795,
      "asia": 2548,
      "favors": 11014,
      "regional": 27819,
      "asian": 2549,
      "sea": 29300,
      "customs": 6567,
      "overlook": 23944,
      "encompasses": 9387,
      "storytelling": 31089,
      "humanannotated": 13884,
      "documentgrounded": 8513,
      "curvature": 6552,
      "underscored": 34178,
      "substitution": 31493,
      "molecules": 22202,
      "textattributed": 32969,
      "aligner": 1743,
      "disadvantages": 8249,
      "succeeded": 31506,
      "highlighter": 13641,
      "multimodality": 22369,
      "heavier": 13491,
      "reliance": 27959,
      "interactively": 15396,
      "highlighted": 13639,
      "vlms": 35377,
      "confirm": 5659,
      "707": 502,
      "mmbench": 20309,
      "degrading": 7384,
      "forcing": 11719,
      "advocates": 1533,
      "amber": 1871,
      "committed": 5000,
      "boundaries": 3843,
      "coefficient": 4870,
      "promotional": 26306,
      "sales": 29058,
      "chatgpt4": 4503,
      "bing": 3737,
      "promotion": 26305,
      "consumergrade": 5819,
      "pooling": 25100,
      "idle": 14033,
      "volunteers": 35390,
      "device": 7989,
      "disconnect": 8262,
      "abruptly": 757,
      "partition": 24359,
      "decentralized": 7224,
      "180": 184,
      "releasing": 27930,
      "ppo": 25357,
      "sustainability": 31785,
      "wideranging": 35588,
      "sustainable": 31786,
      "816": 562,
      "demos": 7598,
      "steady": 31026,
      "dialoguebased": 8022,
      "lowest": 19295,
      "placing": 24986,
      "5th": 456,
      "2nd": 299,
      "aibased": 1625,
      "stress": 31144,
      "confusion": 5675,
      "agitation": 1586,
      "multidimensional": 22275,
      "gathered": 12127,
      "shaping": 29778,
      "conducive": 5580,
      "comprehensible": 5339,
      "visually": 35366,
      "manuscript": 19574,
      "company": 5070,
      "readable": 27296,
      "55": 446,
      "longcontext": 19192,
      "4k": 414,
      "theoretically": 33052,
      "noticed": 23041,
      "multidocument": 22276,
      "kgs": 15799,
      "kg": 15798,
      "intermediary": 15422,
      "adaptable": 1172,
      "assimilate": 2630,
      "528": 437,
      "locating": 19147,
      "log": 19150,
      "odia": 23281,
      "unavailability": 34111,
      "indic": 14684,
      "citizens": 4573,
      "mixtral8x7b": 20276,
      "preprocess": 25502,
      "marking": 19600,
      "breakthrough": 3861,
      "assisted": 2638,
      "forensics": 11727,
      "positioning": 25189,
      "unresolved": 34428,
      "corrections": 6201,
      "solar": 30459,
      "upscaling": 34485,
      "depthwise": 7668,
      "mixtral8x7binstruct": 20277,
      "harvesting": 13469,
      "grapple": 13242,
      "crafting": 6342,
      "outliers": 23741,
      "establishment": 9779,
      "elucidates": 9115,
      "minimizes": 20198,
      "principled": 25932,
      "questioning": 27093,
      "llama12": 18155,
      "mutual": 22471,
      "hinting": 13723,
      "criticism": 6401,
      "nonlinear": 23001,
      "compensation": 5209,
      "definitely": 7367,
      "famous": 10985,
      "afford": 1548,
      "prices": 25906,
      "ablations": 739,
      "languagebased": 16856,
      "safely": 29038,
      "selfdriving": 29418,
      "planner": 24997,
      "navigates": 22593,
      "primer": 25926,
      "operated": 23562,
      "zephyr7bbeta": 35934,
      "interdiscipline": 15404,
      "earth": 8786,
      "34": 335,
      "diversified": 8474,
      "entailed": 9614,
      "synthesize": 31839,
      "subcategories": 31419,
      "veracity": 35201,
      "inaccuracies": 14432,
      "singlehop": 30232,
      "trigger": 33898,
      "exemplified": 10205,
      "englishdominant": 9496,
      "informativeness": 14934,
      "harmlessness": 13449,
      "plant": 25005,
      "evolved": 10078,
      "plants": 25006,
      "engineers": 9472,
      "7bparameter": 546,
      "culminating": 6457,
      "upcoming": 34472,
      "iot": 15637,
      "sensing": 29512,
      "networking": 22706,
      "smart": 30407,
      "household": 13769,
      "robots": 28928,
      "dalle": 6580,
      "views": 35277,
      "security": 29348,
      "tinyllama": 33175,
      "opposite": 23585,
      "laying": 17451,
      "solid": 30465,
      "deepseek": 7353,
      "mixtral": 20273,
      "8x7b": 588,
      "timestep": 33169,
      "claude21": 4640,
      "susceptible": 31782,
      "suspected": 31783,
      "dropped": 8744,
      "associative": 2657,
      "mllms": 20294,
      "economics": 8812,
      "economy": 8813,
      "traffic": 33357,
      "burdens": 3942,
      "adults": 1415,
      "disorder": 8310,
      "credibility": 6373,
      "prefer": 25465,
      "truthfully": 33938,
      "suffering": 31556,
      "rooted": 28972,
      "assigned": 2624,
      "nouns": 23050,
      "proper": 26468,
      "adjectives": 1388,
      "elicit": 9103,
      "hallucinated": 13369,
      "webscale": 35485,
      "textitie": 32983,
      "repeating": 28092,
      "ngrambased": 22902,
      "ultimate": 34096,
      "specialization": 30662,
      "acquires": 1122,
      "principal": 25930,
      "finely": 11278,
      "segmenting": 29371,
      "activating": 1137,
      "285": 288,
      "maybe": 19708,
      "182": 188,
      "notion": 23045,
      "deciphering": 7231,
      "curved": 6554,
      "studys": 31411,
      "infrequent": 14938,
      "querydocument": 27034,
      "phi": 24926,
      "ragbased": 27161,
      "apibased": 2105,
      "fetch": 11081,
      "rlbased": 28905,
      "threshold": 33096,
      "getting": 12734,
      "accomplish": 840,
      "generalist": 12197,
      "specialist": 30660,
      "decreases": 7301,
      "equitable": 9684,
      "quantisation": 26989,
      "hit": 13731,
      "contributor": 6045,
      "incurs": 14663,
      "161": 160,
      "socalled": 30415,
      "checks": 4524,
      "redaction": 27691,
      "trusted": 33933,
      "selfreflective": 29431,
      "deficiencies": 7360,
      "proofs": 26462,
      "487": 406,
      "wider": 35583,
      "pdfs": 24428,
      "penetration": 24449,
      "72": 509,
      "paving": 24420,
      "proxy": 26799,
      "operates": 23563,
      "tunes": 33964,
      "proxies": 26792,
      "3times": 371,
      "replaces": 28104,
      "constants": 5772,
      "roleplaying": 28963,
      "emotionally": 9206,
      "realism": 27314,
      "meticulous": 20112,
      "authentic": 2832,
      "immersive": 14117,
      "300b": 310,
      "cascaded": 4192,
      "comparably": 5094,
      "saw": 29118,
      "selfrewarding": 29434,
      "superhuman": 31642,
      "bottlenecked": 3840,
      "axes": 3047,
      "summaries": 31605,
      "gpt4based": 13129,
      "potent": 25233,
      "blending": 3762,
      "impractical": 14226,
      "collective": 4936,
      "elevating": 9102,
      "fragment": 11819,
      "unicode": 34318,
      "expedite": 10353,
      "discovers": 8273,
      "looking": 19221,
      "fscore": 11942,
      "apt": 2424,
      "discarding": 8255,
      "remained": 27989,
      "8x": 587,
      "japanese": 15699,
      "korean": 15941,
      "streamlined": 31135,
      "semiautoregressive": 29493,
      "treebased": 33879,
      "delving": 7408,
      "corruption": 6237,
      "encapsulated": 9333,
      "sc": 29120,
      "transitions": 33810,
      "aigc": 1631,
      "tampered": 32043,
      "eye": 10810,
      "documentlevel": 8515,
      "laborintensive": 15971,
      "longtext": 19218,
      "inspiring": 15103,
      "proceed": 26047,
      "picking": 24946,
      "orchestration": 23667,
      "mixtrals": 20278,
      "pool": 25099,
      "synergy": 31808,
      "modelsmllms": 22148,
      "continues": 5991,
      "imposes": 14222,
      "comprehensiveness": 5400,
      "fillintheblank": 11167,
      "randomness": 27184,
      "gpt4v": 13135,
      "geminipro": 12144,
      "qwenvlplus": 27140,
      "turkish": 34028,
      "wellresourced": 35528,
      "ul2": 34094,
      "competes": 5214,
      "deleting": 7395,
      "rows": 28997,
      "columns": 4946,
      "phi2": 24929,
      "24gb": 266,
      "40gb": 380,
      "autoregression": 2934,
      "languagemodel": 16860,
      "equipping": 9682,
      "propagation": 26465,
      "offload": 23323,
      "solver": 30502,
      "internally": 15445,
      "adeptness": 1380,
      "linked": 18029,
      "gradually": 13201,
      "imagetext": 14092,
      "inevitable": 14752,
      "adaption": 1218,
      "route": 28988,
      "llava15": 18251,
      "micro": 20154,
      "imposing": 14223,
      "pet": 24915,
      "fever": 11083,
      "permutations": 24869,
      "decompositions": 7291,
      "gpt2small": 12977,
      "124": 94,
      "lutbased": 19331,
      "256": 274,
      "anticipatory": 2095,
      "anticipating": 2094,
      "normative": 23022,
      "mistral7binstruct": 20240,
      "seldom": 29374,
      "threads": 33088,
      "surged": 31730,
      "undisclosed": 34295,
      "cutoff": 6569,
      "reflections": 27796,
      "ownership": 23960,
      "distill": 8333,
      "therapeutic": 33061,
      "wish": 35623,
      "listening": 18034,
      "client": 4652,
      "guess": 13338,
      "holdout": 13737,
      "figure": 11160,
      "outcome": 23734,
      "accepted": 814,
      "diminishing": 8207,
      "userprovided": 34682,
      "outofvocabulary": 23759,
      "e2e": 8769,
      "messaging": 19855,
      "moebased": 22199,
      "determined": 7900,
      "simulate": 30184,
      "humancomputer": 13892,
      "voting": 35392,
      "cautious": 4258,
      "democratic": 7422,
      "marginalized": 19588,
      "turning": 34030,
      "questiongeneration": 27092,
      "lays": 17456,
      "discussing": 8298,
      "lighter": 17834,
      "systemlevel": 31885,
      "guardrails": 13334,
      "safeguarding": 29036,
      "guard": 13333,
      "embrace": 9152,
      "equivalently": 9688,
      "cached": 3959,
      "enriching": 9594,
      "certified": 4285,
      "ticket": 33102,
      "guaranteed": 13330,
      "malaysian": 19520,
      "recallk": 27472,
      "malay": 19519,
      "humanmachine": 13914,
      "ev": 9812,
      "compile": 5238,
      "deductive": 7309,
      "satisfactory": 29107,
      "supportive": 31716,
      "incorporation": 14580,
      "inaccurate": 14434,
      "412": 384,
      "instantiation": 15117,
      "amplifying": 1897,
      "drawn": 8729,
      "november": 23128,
      "aya": 3049,
      "ift": 14035,
      "worked": 35799,
      "513": 432,
      "114": 78,
      "toolaugmented": 33262,
      "selfevaluation": 29419,
      "affects": 1544,
      "pink": 24955,
      "elephant": 9101,
      "critiques": 6402,
      "reshape": 28386,
      "threats": 33091,
      "malicious": 19522,
      "actors": 1153,
      "surveyed": 31777,
      "round": 28986,
      "bots": 3837,
      "persona": 24879,
      "amharic": 1880,
      "sought": 30541,
      "examined": 10105,
      "newer": 22867,
      "minor": 20204,
      "contingent": 5972,
      "viewpoint": 35275,
      "treats": 33875,
      "arm": 2506,
      "incurred": 14660,
      "triple": 33913,
      "embeddingbased": 9136,
      "welladopted": 35519,
      "internetscale": 15448,
      "quantizes": 27016,
      "highprecision": 13681,
      "deltas": 7405,
      "soup": 30545,
      "leaves": 17680,
      "verbalizers": 35204,
      "toy": 33320,
      "mimics": 20173,
      "pivot": 24975,
      "region": 27818,
      "consequences": 5694,
      "tasked": 32215,
      "intrinsically": 15495,
      "modelspecific": 22149,
      "variances": 35039,
      "opening": 23466,
      "11m": 86,
      "haystack": 13475,
      "miss": 20221,
      "augmentations": 2809,
      "106": 60,
      "achievement": 1022,
      "longest": 19203,
      "crisis": 6375,
      "emergency": 9180,
      "911": 595,
      "notify": 23043,
      "incident": 14444,
      "overwhelmed": 23956,
      "hardwarefriendly": 13436,
      "silicon": 30091,
      "codesign": 4857,
      "normalization": 23017,
      "parallelization": 24174,
      "anchored": 2005,
      "president": 25613,
      "outdated": 23737,
      "updated": 34475,
      "rerunning": 28282,
      "roleoriented": 28961,
      "secret": 29338,
      "geared": 12135,
      "usecases": 34581,
      "chapters": 4420,
      "llemma": 18258,
      "finishing": 11560,
      "executor": 10202,
      "codebased": 4834,
      "10k": 63,
      "lifelong": 17822,
      "openbook": 23448,
      "retention": 28722,
      "sensory": 29526,
      "chance": 4405,
      "solvable": 30486,
      "6000": 460,
      "llama2chat70b": 18221,
      "singletask": 30234,
      "anchors": 2007,
      "02": 4,
      "judgement": 15722,
      "gpt4turbo": 13134,
      "selfdistillation": 29417,
      "ignoring": 14040,
      "forgotten": 11737,
      "wellunderstood": 35532,
      "edits": 8833,
      "xxl": 35887,
      "analogous": 1903,
      "v2": 34984,
      "needle": 22649,
      "hardwareaware": 13434,
      "speculation": 30776,
      "temperatures": 32688,
      "maximizes": 19703,
      "vicuna33b": 35259,
      "56": 448,
      "behavioural": 3332,
      "illuminate": 14044,
      "strands": 31095,
      "semeval2024": 29491,
      "relatedness": 27860,
      "official": 23319,
      "1st": 204,
      "collaborating": 4903,
      "topperforming": 33296,
      "arabiccentric": 2426,
      "relu": 27967,
      "gelu": 12136,
      "substituting": 31492,
      "curves": 6555,
      "granting": 13214,
      "trick": 33894,
      "beam": 3299,
      "hoping": 13760,
      "likert": 17903,
      "profile": 26187,
      "questionnaire": 27094,
      "populations": 25148,
      "instructed": 15134,
      "diminishes": 8202,
      "coherently": 4897,
      "da": 6576,
      "prolific": 26262,
      "spending": 30805,
      "negates": 22655,
      "concluding": 5555,
      "echo": 8805,
      "repeat": 28089,
      "maximally": 19700,
      "07": 14,
      "drafts": 8714,
      "peerreview": 24433,
      "manuscripts": 19575,
      "onerous": 23343,
      "summarizing": 31631,
      "mezo": 20152,
      "zerothorder": 36000,
      "zo": 36005,
      "llama30b": 18224,
      "chimera": 4537,
      "shortrange": 29827,
      "languagecentric": 16858,
      "exhaustive": 10209,
      "equips": 9683,
      "guarantees": 13332,
      "highconfidence": 13590,
      "tight": 33105,
      "rest": 28526,
      "competencies": 5212,
      "structurally": 31211,
      "cumbersome": 6464,
      "videos": 35268,
      "intellectual": 15347,
      "patent": 24396,
      "noticeable": 23039,
      "situated": 30237,
      "steer": 31027,
      "deactivating": 7206,
      "clock": 4670,
      "unaligned": 34104,
      "hint": 13722,
      "gisting": 12736,
      "economical": 8811,
      "continuations": 5984,
      "ece": 8804,
      "advise": 1529,
      "newton": 22890,
      "participate": 24328,
      "disentangle": 8305,
      "attributevalue": 2766,
      "adversely": 1526,
      "deviate": 7985,
      "publishing": 26872,
      "insert": 15056,
      "formation": 11756,
      "enrichment": 9595,
      "ontological": 23375,
      "locates": 19146,
      "finetuningbased": 11558,
      "ct": 6453,
      "linking": 18030,
      "crossencoder": 6408,
      "accelerated": 795,
      "pt": 26825,
      "multiquery": 22435,
      "querybased": 27033,
      "stays": 31025,
      "endpoints": 9430,
      "premature": 25492,
      "flows": 11629,
      "patching": 24395,
      "circuits": 4569,
      "counts": 6307,
      "instructionresponse": 15239,
      "24times": 268,
      "anchoring": 2006,
      "timelines": 33154,
      "nextstep": 22892,
      "codetocode": 4860,
      "twophase": 34041,
      "warmup": 35422,
      "unveils": 34469,
      "lay": 17422,
      "groundwork": 13295,
      "monotonic": 22213,
      "meaningless": 19727,
      "punctuation": 26878,
      "visualizations": 35361,
      "conveying": 6118,
      "indispensable": 14711,
      "programbased": 26196,
      "termination": 32733,
      "roleplay": 28962,
      "nonreproducible": 23005,
      "rtx": 28999,
      "4090": 377,
      "curating": 6473,
      "conjectures": 5677,
      "5200": 434,
      "postedit": 25223,
      "winrate": 35619,
      "elementary": 9099,
      "typologically": 34084,
      "warrant": 35423,
      "folds": 11674,
      "lagging": 16015,
      "divideandconquer": 8487,
      "manageable": 19527,
      "aggregates": 1578,
      "counting": 6304,
      "viability": 35240,
      "conducts": 5646,
      "spite": 30809,
      "evoking": 10072,
      "progresses": 26230,
      "chatbased": 4447,
      "correction": 6198,
      "2based": 294,
      "f05": 10813,
      "invoking": 15623,
      "analyst": 1980,
      "kv": 15947,
      "row": 28996,
      "deduplicating": 7313,
      "vllm": 35374,
      "essay": 9752,
      "aes": 1534,
      "extraordinary": 10785,
      "holistically": 13744,
      "tactics": 32008,
      "blueprint": 3799,
      "divide": 8486,
      "unverified": 34470,
      "flagging": 11584,
      "caregivers": 4181,
      "fms": 11641,
      "fm": 11639,
      "gathering": 12128,
      "returning": 28782,
      "carefullydesigned": 4180,
      "recitation": 27634,
      "selective": 29398,
      "manage": 19526,
      "permutation": 24868,
      "tokenized": 33210,
      "complemented": 5245,
      "mp": 22242,
      "jetson": 15705,
      "interlinear": 15421,
      "ultra": 34098,
      "documentation": 8510,
      "gemma": 12145,
      "stateofthe": 30917,
      "webcrawled": 35481,
      "lowerresourced": 19293,
      "european": 9810,
      "confronts": 5672,
      "diverging": 8409,
      "refines": 27784,
      "negating": 22656,
      "wizardlms": 35627,
      "aidriven": 1629,
      "opt67b": 23610,
      "rogue": 28951,
      "offset": 23325,
      "creativity": 6372,
      "devising": 7994,
      "512": 431,
      "576": 451,
      "weather": 35469,
      "emulating": 9277,
      "aging": 1585,
      "brain": 3850,
      "disorders": 8311,
      "loses": 19238,
      "tech": 32603,
      "lessresourced": 17724,
      "lrls": 19321,
      "lrl": 19320,
      "skew": 30306,
      "sociolinguistic": 30442,
      "fullmodel": 11947,
      "justintime": 15737,
      "compilation": 5237,
      "12x": 105,
      "vram": 35398,
      "streaming": 31133,
      "15k": 153,
      "resorts": 28407,
      "fiction": 11129,
      "solvers": 30503,
      "specializes": 30677,
      "confidencebased": 5651,
      "corrects": 6210,
      "directives": 8231,
      "grid": 13283,
      "unidirectional": 34319,
      "longbench": 19191,
      "builtin": 3939,
      "13000": 111,
      "stars": 30896,
      "twoparty": 34040,
      "audit": 2784,
      "debates": 7212,
      "reevaluate": 27768,
      "llmsbased": 19051,
      "exploited": 10557,
      "essentially": 9764,
      "unlikelihood": 34407,
      "averagely": 3029,
      "sacrebleu": 29023,
      "bleurt": 3770,
      "react": 27289,
      "selftraining": 29443,
      "accumulated": 854,
      "mistral7binstructv02": 20241,
      "alfworld": 1698,
      "1shot": 203,
      "webshop": 35486,
      "deviation": 7987,
      "counter": 6290,
      "randomized": 27179,
      "compromise": 5436,
      "adamw": 1161,
      "345": 336,
      "345m": 337,
      "326": 322,
      "526": 436,
      "sst2": 30845,
      "specifying": 30768,
      "summarized": 31630,
      "imdb": 14105,
      "umls": 34099,
      "28b": 290,
      "sentencet5": 29563,
      "confirmed": 5661,
      "noteworthy": 23037,
      "300": 309,
      "763": 519,
      "iti": 15693,
      "constitutes": 5776,
      "shifted": 29805,
      "multitoken": 22458,
      "testsets": 32810,
      "invasive": 15565,
      "kullbackleibler": 15945,
      "stance": 30864,
      "ukraine": 34092,
      "war": 35420,
      "segment": 29369,
      "mtl": 22263,
      "avg": 3033,
      "vulnerability": 35404,
      "quantifies": 26985,
      "adeptly": 1379,
      "rewritten": 28870,
      "gai": 12054,
      "acceptance": 812,
      "individualized": 14720,
      "extant": 10645,
      "efficacious": 8984,
      "englishlanguage": 9498,
      "administrative": 1398,
      "mlps": 20307,
      "erases": 9704,
      "neural machine": 22731,
      "machine translation": 19355,
      "gpt2 bert": 12875,
      "demonstrate effectiveness": 7445,
      "effectiveness using": 8970,
      "using pretrained": 34878,
      "pretrained language": 25653,
      "language models": 16231,
      "models lms": 21668,
      "lms various": 19122,
      "various natural": 35124,
      "natural language": 22507,
      "language processing": 16774,
      "processing tasks": 26126,
      "catastrophic forgetting": 4213,
      "tasks work": 32553,
      "work introduce": 35723,
      "training framework": 33523,
      "pretrained lms": 25708,
      "translation nmt": 33839,
      "nmt model": 22978,
      "previous pretrained": 25872,
      "pretrained knowledge": 25652,
      "bleu score": 3768,
      "language pair": 16766,
      "surpasses previous": 31749,
      "previous stateoftheart": 25881,
      "score large": 29272,
      "base model": 3126,
      "model significantly": 20783,
      "significantly improves": 30060,
      "improves stateoftheart": 14394,
      "stateoftheart transformer": 31003,
      "model bleu": 20402,
      "code model": 4779,
      "models large": 21420,
      "large language": 16959,
      "models range": 21885,
      "dataset biases": 6944,
      "generative capabilities": 12658,
      "release gpt2": 27908,
      "gpt2 language": 12908,
      "language model": 16114,
      "time model": 33136,
      "model sizes": 20796,
      "research provides": 28352,
      "masked language": 19610,
      "models mlms": 21734,
      "require finetuning": 28217,
      "nlp tasks": 22948,
      "tasks instead": 32377,
      "autoregressive language": 2942,
      "models like": 21447,
      "like gpt2": 17867,
      "gpt2 variety": 12966,
      "variety tasks": 35071,
      "rescoring asr": 28284,
      "stateoftheart baselines": 30924,
      "lowresource translation": 19318,
      "translation pairs": 33841,
      "domain adaptation": 8552,
      "linguistic acceptability": 18004,
      "scores gpt2": 29278,
      "10 points": 30,
      "single inference": 30207,
      "use growing": 34535,
      "growing number": 13316,
      "number pretrained": 23158,
      "use single": 34569,
      "crosslingual model": 6416,
      "multiple languages": 22397,
      "constrained text": 5781,
      "text generation": 32865,
      "generation challenge": 12469,
      "generative commonsense": 12659,
      "commonsense reasoning": 5039,
      "reasoning recently": 27447,
      "recently largescale": 27610,
      "largescale pretrained": 17373,
      "models demonstrated": 21134,
      "demonstrated impressive": 7526,
      "impressive performance": 14241,
      "benchmark datasets": 3373,
      "remains challenging": 27993,
      "challenging paper": 4390,
      "paper present": 24085,
      "generation task": 12611,
      "benchmark dataset": 3371,
      "ability generative": 678,
      "reasoning given": 27410,
      "task generate": 32130,
      "generate coherent": 12267,
      "using concepts": 34755,
      "task challenging": 32092,
      "commonsense knowledge": 5036,
      "compositional generalization": 5327,
      "generalization ability": 12205,
      "dataset constructed": 6961,
      "experiments large": 10456,
      "large gap": 16949,
      "stateoftheart text": 30998,
      "generation models": 12552,
      "models t5": 22041,
      "human performance": 13851,
      "performance furthermore": 24604,
      "furthermore demonstrate": 11992,
      "demonstrate learned": 7470,
      "reasoning capability": 27390,
      "improve downstream": 14262,
      "downstream tasks": 8687,
      "additional context": 1253,
      "better text": 3631,
      "text understanding": 32959,
      "understanding recent": 34266,
      "recent progress": 27541,
      "progress nlp": 26222,
      "nlp witnessed": 22968,
      "models gpt": 21323,
      "bert xlnet": 3537,
      "based transformer": 3232,
      "et al": 9789,
      "end tasks": 9419,
      "tasks models": 32421,
      "models achieved": 20947,
      "achieved stateoftheart": 1017,
      "stateoftheart results": 30983,
      "sufficient number": 31562,
      "number layers": 23150,
      "large pretraining": 17271,
      "pretraining data": 25791,
      "data tasks": 6889,
      "tasks require": 32482,
      "gap pretrained": 12104,
      "pretrained models": 25716,
      "models human": 21362,
      "al 2018": 1681,
      "syntactic structure": 31825,
      "structure model": 31215,
      "model supervised": 20814,
      "semantic knowledge": 29459,
      "coreference information": 6157,
      "information existing": 14865,
      "existing model": 10298,
      "model improve": 20571,
      "improve performance": 14282,
      "performance complex": 24555,
      "complex problems": 5283,
      "al 2016": 1680,
      "task model": 32159,
      "model trained": 20834,
      "trained scratch": 33423,
      "auxiliary supervision": 2960,
      "outperforms largest": 23828,
      "largest gpt2": 17393,
      "gpt2 model": 12916,
      "setting new": 29724,
      "new stateoftheart": 22848,
      "tiny fraction": 33172,
      "fraction parameters": 11818,
      "parameters compared": 24232,
      "compared gpt2": 5135,
      "conduct thorough": 5625,
      "thorough analysis": 33071,
      "analysis different": 1918,
      "different variants": 8157,
      "model architectures": 20376,
      "future directions": 12031,
      "similar techniques": 30119,
      "learning semantic": 17652,
      "text modeling": 32911,
      "requires commonsense": 28248,
      "knowledge world": 15923,
      "exploring various": 10623,
      "various knowledge": 35103,
      "representations previous": 28172,
      "previous work": 25891,
      "work focused": 35712,
      "methods fail": 20034,
      "supervised setting": 31689,
      "large pretrained": 17259,
      "models led": 21443,
      "improved results": 14318,
      "results natural": 28649,
      "language understanding": 16838,
      "understanding tasks": 34274,
      "difficult problem": 8172,
      "learning model": 17607,
      "text create": 32839,
      "create training": 6356,
      "training set": 33609,
      "events large": 10048,
      "large corpus": 16938,
      "extend idea": 10652,
      "models machine": 21695,
      "machine learning": 19343,
      "learning tasks": 17661,
      "multilayer transformer": 22293,
      "models high": 21353,
      "high accuracy": 13549,
      "outperform models": 23780,
      "models similar": 21977,
      "similar size": 30118,
      "degree models": 7389,
      "models larger": 21431,
      "larger size": 17337,
      "size trained": 30287,
      "trained using": 33433,
      "computational budget": 5454,
      "key observation": 15780,
      "alternative method": 1854,
      "solving problems": 30516,
      "large vocabulary": 17300,
      "parameters language": 24258,
      "model recently": 20745,
      "neural language": 22724,
      "models trained": 22065,
      "unstructured text": 34451,
      "retrieve knowledge": 28768,
      "knowledge using": 15919,
      "using natural": 34849,
      "language queries": 16809,
      "short paper": 29813,
      "practical utility": 25374,
      "approach finetuning": 2283,
      "finetuning pretrained": 11486,
      "models answer": 20980,
      "answer questions": 2051,
      "access external": 816,
      "context knowledge": 5894,
      "knowledge approach": 15813,
      "model size": 20788,
      "performs competitively": 24848,
      "external knowledge": 10728,
      "knowledge source": 15908,
      "answering questions": 2071,
      "facilitate reproducibility": 10842,
      "future work": 12052,
      "release code": 27901,
      "code trained": 4824,
      "trained models": 33413,
      "data augmentation": 6605,
      "augmentation using": 2808,
      "pretrained transformer": 25756,
      "transformer models": 33731,
      "models language": 21415,
      "model based": 20386,
      "based pretrained": 3208,
      "models bert": 21015,
      "significant gains": 29982,
      "gains different": 12071,
      "different nlp": 8112,
      "tasks paper": 32440,
      "paper study": 24137,
      "study different": 31318,
      "different types": 8154,
      "transformer based": 33706,
      "models autoregressive": 21000,
      "autoregressive models": 2953,
      "models gpt2": 21324,
      "autoencoder models": 2843,
      "seq2seq models": 29592,
      "models bart": 21005,
      "class labels": 4585,
      "labels text": 15968,
      "text sequences": 32938,
      "simple effective": 30143,
      "effective way": 8906,
      "models data": 21121,
      "classification benchmarks": 4593,
      "pretrained seq2seq": 25746,
      "seq2seq model": 29591,
      "model outperforms": 20670,
      "augmentation methods": 2804,
      "lowresource setting": 19316,
      "explore different": 10581,
      "different pretrained": 8123,
      "pretrained model": 25711,
      "based data": 3149,
      "data diversity": 6683,
      "generative pretraining": 12701,
      "generation evaluation": 12494,
      "automatic generation": 2882,
      "past years": 24392,
      "instruction generation": 15169,
      "generation given": 12512,
      "generation module": 12557,
      "generative pretrained": 12685,
      "model gpt2": 20554,
      "gpt2 finetuned": 12891,
      "finetuned large": 11324,
      "allows users": 1818,
      "quality generated": 26962,
      "accessed online": 829,
      "information seeking": 14911,
      "reading comprehension": 27305,
      "marco datasets": 19582,
      "20 test": 213,
      "ranking methods": 27232,
      "methods include": 20048,
      "traditional retrieval": 33351,
      "retrieval based": 28738,
      "neural models": 22737,
      "models knowledge": 21410,
      "knowledge enhanced": 15843,
      "bertbased neural": 3542,
      "reranking methods": 28281,
      "methods employed": 20025,
      "query expansion": 27026,
      "expansion generative": 10345,
      "generative language": 12661,
      "models conversational": 21111,
      "gpt2 results": 12947,
      "automatic systems": 2898,
      "systems using": 31924,
      "relative improvement": 27880,
      "conversational question": 6102,
      "architectures pretrained": 2470,
      "models paper": 21790,
      "paper presents": 24095,
      "presents empirical": 25582,
      "empirical study": 9240,
      "study conversational": 31312,
      "models plms": 21819,
      "independence assumption": 14669,
      "maximum likelihood": 19706,
      "likelihood estimation": 17899,
      "benchmarks taskoriented": 3473,
      "taskoriented dialogue": 32221,
      "dialogue systems": 8020,
      "systems evaluate": 31896,
      "models using": 22101,
      "using data": 34762,
      "different numbers": 8115,
      "numbers parameters": 23175,
      "parameters demonstrate": 24237,
      "demonstrate recent": 7486,
      "texttotext transfer": 33016,
      "transfer transformer": 33683,
      "transformer t5": 33741,
      "achieves best": 1031,
      "best results": 3577,
      "fewer parameters": 11089,
      "transformer architectures": 33703,
      "dynamic evaluation": 8758,
      "evaluation language": 9963,
      "language use": 16851,
      "new challenge": 22785,
      "challenge task": 4330,
      "task dataset": 32102,
      "understanding models": 34253,
      "models given": 21318,
      "model generate": 20543,
      "generate helpful": 12281,
      "language evaluation": 16070,
      "evaluation framework": 9948,
      "fundamental aspect": 11972,
      "aspect human": 2564,
      "human language": 13837,
      "understanding ability": 34207,
      "ability use": 726,
      "use language": 34540,
      "empirical results": 9232,
      "models struggle": 22016,
      "multibillion parameter": 22270,
      "parameter models": 24191,
      "models finetuned": 21274,
      "indomain training": 14726,
      "training examples": 33516,
      "best model": 3563,
      "model finetuned": 20524,
      "finetuned t5": 11355,
      "cases larger": 4203,
      "gpt3 model": 13002,
      "model does": 20472,
      "low performance": 19270,
      "generative setting": 12704,
      "data augmented": 6615,
      "relation extraction": 27864,
      "realworld relation": 27344,
      "extraction tasks": 10774,
      "tasks challenging": 32260,
      "limited training": 17970,
      "training data": 33466,
      "class imbalance": 4584,
      "imbalance issues": 14100,
      "issues work": 15676,
      "work present": 35747,
      "present data": 25522,
      "simple method": 30157,
      "method augment": 19882,
      "augment training": 2793,
      "finetuning gpt2": 11411,
      "gpt2 generate": 12893,
      "types generated": 34062,
      "data used": 6901,
      "dataset train": 7046,
      "series experiments": 29635,
      "method leads": 19941,
      "improvements 11": 14353,
      "11 f1": 70,
      "f1 score": 10817,
      "score points": 29274,
      "strong baseline": 31162,
      "achieves new": 1052,
      "new state": 22846,
      "state art": 30903,
      "widely used": 35576,
      "used biomedical": 34588,
      "datasets surpassing": 7177,
      "surpassing previous": 31759,
      "previous best": 25867,
      "f1 points": 10815,
      "points average": 25070,
      "italian language": 15678,
      "years pretrained": 35895,
      "pretrained neural": 25738,
      "neural architectures": 22721,
      "improvements nlp": 14361,
      "tasks generative": 32348,
      "models available": 21001,
      "built using": 3937,
      "using gpt2": 34789,
      "gpt2 architecture": 12870,
      "provide thorough": 26731,
      "humanbased evaluation": 13890,
      "evaluation automatic": 9922,
      "human evaluation": 13805,
      "evaluation performed": 9986,
      "sentence completion": 29529,
      "original human": 23707,
      "human texts": 13871,
      "simpler language": 30167,
      "tasks study": 32516,
      "pretrain finetune": 25622,
      "tasks experiments": 32323,
      "experiments indicate": 10450,
      "based models": 3197,
      "models outperform": 21782,
      "datatotext generation": 7195,
      "based pretraining": 3210,
      "bert gpt2": 3509,
      "t5 pretraining": 31961,
      "leads better": 17489,
      "better generalization": 3605,
      "generalization evidenced": 12214,
      "large improvements": 16955,
      "improvements outofdomain": 14363,
      "outofdomain test": 23754,
      "test sets": 32788,
      "hope work": 13757,
      "work serves": 35778,
      "baseline future": 3245,
      "future research": 12040,
      "transfer learning": 33676,
      "tasks common": 32267,
      "common sense": 5012,
      "world knowledge": 35834,
      "knowledge injection": 15865,
      "pretrained transformers": 25771,
      "success neural": 31521,
      "lms bert": 19072,
      "variety language": 35062,
      "tasks recent": 32474,
      "recent work": 27567,
      "structured knowledge": 31222,
      "knowledge external": 15851,
      "external resources": 10733,
      "resources models": 28440,
      "models hand": 21346,
      "pretraining training": 25848,
      "training scratch": 33606,
      "knowledge primary": 15891,
      "computationally expensive": 5486,
      "knowledge work": 15920,
      "work investigate": 35727,
      "investigate models": 15589,
      "knowledge bert": 15821,
      "respectively using": 28465,
      "overall results": 23913,
      "glue benchmark": 12802,
      "analysis reveals": 1959,
      "models substantially": 22022,
      "substantially outperform": 31484,
      "inference tasks": 14814,
      "knowledge explicitly": 15850,
      "code experiments": 4749,
      "open sourced": 23427,
      "evaluation stateoftheart": 10013,
      "stateoftheart nlp": 30966,
      "deep learning": 7323,
      "learning architectures": 17540,
      "reasoning task": 27455,
      "task paper": 32171,
      "paper investigate": 24066,
      "investigate commonsense": 15579,
      "inference task": 14813,
      "task competition": 32095,
      "stateoftheart deep": 30929,
      "manually curated": 19568,
      "language inference": 16094,
      "make sense": 19482,
      "sense make": 29507,
      "compare performance": 5111,
      "performance language": 24640,
      "propose method": 26528,
      "method inspired": 19935,
      "questionanswering tasks": 27089,
      "classification problem": 4604,
      "multiple choice": 22382,
      "choice question": 4552,
      "boost performance": 3817,
      "performance experimental": 24591,
      "experimental results": 10389,
      "significantly better": 30035,
      "better baseline": 3594,
      "does make": 8533,
      "competitive results": 5229,
      "results result": 28671,
      "future researches": 12050,
      "powerful generative": 25338,
      "generative model": 12672,
      "model language": 20600,
      "language gpt2": 16092,
      "fewshot generative": 11105,
      "rewriting aims": 28868,
      "existing information": 10275,
      "information retrieval": 14908,
      "retrieval systems": 28756,
      "systems paper": 31911,
      "presents fewshot": 25584,
      "generative approach": 12649,
      "develop methods": 7917,
      "methods based": 20009,
      "based rules": 3219,
      "selfsupervised learning": 29437,
      "learning generate": 17572,
      "weak supervision": 35456,
      "supervision data": 31696,
      "data using": 6906,
      "using large": 34807,
      "large amounts": 16927,
      "ad hoc": 1158,
      "finetune gpt2": 11283,
      "weakly supervised": 35461,
      "stateoftheart ranking": 30981,
      "accuracy 12": 858,
      "using limited": 34818,
      "zeroshot learning": 35983,
      "learning setting": 17653,
      "stateoftheart systems": 30995,
      "capture context": 4147,
      "model pretraining": 20723,
      "pretraining knowledge": 25804,
      "knowledge pretrained": 15887,
      "recent research": 27549,
      "grasp human": 13244,
      "human knowledge": 13835,
      "transformer architecture": 33702,
      "explicit knowledge": 10546,
      "external storage": 10735,
      "semantic information": 29458,
      "input transformer": 15036,
      "prediction task": 25437,
      "task experiments": 32119,
      "pretraining significantly": 25839,
      "transformer parameters": 33738,
      "observe improved": 23229,
      "language modeling": 16214,
      "factual correctness": 10880,
      "knowledge probing": 15892,
      "probing tasks": 25984,
      "hidden representations": 13537,
      "dropin replacement": 8742,
      "gpt2 models": 12925,
      "models significantly": 21974,
      "significantly improving": 30063,
      "improving downstream": 14407,
      "tasks like": 32399,
      "training deep": 33498,
      "subword units": 31502,
      "morphologically rich": 22222,
      "recently deep": 27587,
      "models proven": 21871,
      "particularly powerful": 24352,
      "powerful language": 25340,
      "modeling tasks": 20908,
      "high complexity": 13555,
      "complexity makes": 5304,
      "makes difficult": 19490,
      "single pass": 30219,
      "recent studies": 27552,
      "knowledge neural": 15884,
      "neural network": 22738,
      "network language": 22692,
      "models lm": 21667,
      "using neural": 34853,
      "neural text": 22760,
      "generation based": 12462,
      "pretrain gpt2": 25623,
      "gpt2 transformer": 12960,
      "general text": 12188,
      "text corpus": 32836,
      "task data": 32101,
      "language propose": 16808,
      "propose new": 26536,
      "new method": 22820,
      "method called": 19887,
      "text augmentation": 32818,
      "generated text": 12392,
      "methods significantly": 20094,
      "significantly improve": 30055,
      "greatly reducing": 13277,
      "size memory": 30263,
      "memory requirements": 19826,
      "finally demonstrate": 11190,
      "using generative": 34785,
      "models work": 22133,
      "work demonstrates": 35692,
      "language transformers": 16836,
      "learning natural": 17619,
      "language skills": 16822,
      "generate meaningful": 12301,
      "finetuning transformer": 11551,
      "game notation": 12077,
      "training steps": 33623,
      "openais generative": 23438,
      "transformer gpt2": 33720,
      "million parameters": 20165,
      "parameters finetuned": 24245,
      "plausible strategies": 25015,
      "novel model": 23100,
      "model demonstrates": 20457,
      "provides novel": 26760,
      "novel method": 23095,
      "work build": 35674,
      "models synthetic": 22037,
      "models shown": 21969,
      "shown accurately": 29869,
      "accurately reflect": 939,
      "human beings": 13798,
      "used train": 34631,
      "train models": 33368,
      "models models": 21736,
      "traditional methods": 33348,
      "model using": 20852,
      "using text": 34928,
      "text generated": 32861,
      "ground truth": 13285,
      "using model": 34843,
      "model substantially": 20811,
      "similar human": 30105,
      "latent representation": 17408,
      "learning models": 17608,
      "models text": 22057,
      "recent years": 27575,
      "fields natural": 11157,
      "processing nlp": 26114,
      "retrieval ir": 28742,
      "recurrent neural": 27681,
      "neural networks": 22752,
      "networks rnns": 22715,
      "gated recurrent": 12125,
      "long shortterm": 19182,
      "shortterm memory": 29829,
      "bidirectional encoder": 3692,
      "encoder representations": 9354,
      "representations transformers": 28175,
      "transformers bert": 33777,
      "deep neural": 7337,
      "world applications": 35833,
      "small model": 30358,
      "size low": 30262,
      "response times": 28482,
      "low computational": 19266,
      "computational power": 5475,
      "pruning quantization": 26815,
      "knowledge distillation": 15830,
      "parameter sharing": 24197,
      "models enable": 21195,
      "critical need": 6391,
      "efficient small": 9058,
      "small models": 30359,
      "recently published": 27616,
      "published work": 26871,
      "work deep": 35686,
      "nlp community": 22926,
      "coherent story": 4895,
      "generative models": 12677,
      "models unsupervised": 22097,
      "study large": 31352,
      "large generative": 16950,
      "ability generate": 674,
      "generate text": 12331,
      "tasks finetuning": 32336,
      "finetuning work": 11555,
      "classifiers trained": 4628,
      "human machinegenerated": 13845,
      "machinegenerated text": 19371,
      "quality able": 26939,
      "able detect": 743,
      "low quality": 19271,
      "quality content": 26948,
      "training enables": 33509,
      "conduct extensive": 5605,
      "qualitative quantitative": 26934,
      "quantitative analysis": 26991,
      "articles making": 2524,
      "study conducted": 31309,
      "evaluation pretrained": 9992,
      "models automatic": 20998,
      "given question": 12764,
      "desired answer": 7762,
      "previous works": 25895,
      "word embeddings": 35637,
      "semantic features": 29456,
      "features extracted": 11032,
      "features manually": 11035,
      "datasets use": 7185,
      "use pretrained": 34560,
      "pretrained embeddings": 25639,
      "models elmo": 21186,
      "elmo bert": 9113,
      "bert gpt": 3506,
      "gpt gpt2": 12848,
      "gpt2 assess": 12871,
      "task train": 32201,
      "train single": 33374,
      "cosine similarity": 6241,
      "models compare": 21071,
      "models previous": 21850,
      "dataset work": 7056,
      "outperformed models": 23794,
      "models conclude": 21086,
      "conclude possible": 5552,
      "sequence generation": 29595,
      "generation largescale": 12538,
      "largescale language": 17358,
      "lms able": 19068,
      "language generate": 16081,
      "generate realistic": 12316,
      "realistic text": 27318,
      "datasets used": 7186,
      "used training": 34632,
      "training large": 33542,
      "large lms": 17226,
      "usually contain": 34946,
      "efficient method": 9048,
      "method using": 19983,
      "using smaller": 34917,
      "smaller lms": 30380,
      "guide generation": 13345,
      "generation large": 12533,
      "lms make": 19098,
      "method achieving": 19871,
      "30 times": 308,
      "times faster": 33159,
      "additionally training": 1304,
      "generate new": 12307,
      "new capability": 22784,
      "controllable generation": 6058,
      "generation methods": 12547,
      "15b parameters": 152,
      "parameters significantly": 24289,
      "linguistic quality": 18020,
      "models maintaining": 21696,
      "fast generation": 10992,
      "generation speed": 12603,
      "advanced neural": 1437,
      "previous research": 25874,
      "potential abuse": 25235,
      "models assessing": 20990,
      "demonstrates significant": 7567,
      "significant improvement": 29989,
      "gpt2 generating": 12896,
      "generating text": 12445,
      "represents significant": 28197,
      "significant risk": 30021,
      "social norms": 30429,
      "public policy": 26842,
      "disinformation propaganda": 8309,
      "current limitations": 6507,
      "limitations language": 17922,
      "models need": 21748,
      "current approaches": 6482,
      "tradeoff language": 33336,
      "models including": 21377,
      "length efficient": 17705,
      "efficient attention": 9027,
      "openended text": 23462,
      "generation output": 12565,
      "like gpt23": 17868,
      "finetuning dataset": 11385,
      "improve prediction": 14289,
      "scaling model": 29175,
      "size efficiently": 30247,
      "poor performance": 25104,
      "performance scaling": 24746,
      "extend context": 10648,
      "context entire": 5888,
      "entire training": 9627,
      "training dataset": 33495,
      "trained data": 33388,
      "multilabel classification": 22291,
      "paper focus": 24052,
      "strong models": 31183,
      "models suffer": 22027,
      "bias propose": 3661,
      "propose simple": 26565,
      "effective augmentation": 8861,
      "augmentation framework": 2799,
      "takes advantage": 32032,
      "advantage pretrained": 1499,
      "pretrained gpt2": 25649,
      "perturbations input": 24911,
      "input texts": 15033,
      "augment existing": 2790,
      "existing training": 10320,
      "result present": 28544,
      "present substantial": 25557,
      "substantial improvements": 31467,
      "baseline models": 3255,
      "models contributions": 21109,
      "introduce new": 15518,
      "neural toxic": 22762,
      "toxic degeneration": 33309,
      "models pretrained": 21839,
      "lms prone": 19104,
      "prone generating": 26454,
      "toxic language": 33310,
      "safe deployment": 29029,
      "investigate extent": 15582,
      "extent pretrained": 10725,
      "lms prompted": 19102,
      "prompted generate": 26369,
      "controllable text": 6060,
      "generation algorithms": 12457,
      "preventing toxic": 25860,
      "naturally occurring": 22585,
      "corpus english": 6181,
      "web text": 35479,
      "toxic text": 33314,
      "methods data": 20016,
      "adaptive pretraining": 1221,
      "data effective": 6686,
      "text corpora": 32835,
      "corpora used": 6171,
      "used pretrain": 34616,
      "lms including": 19090,
      "including gpt2": 14482,
      "gpt2 radford": 12942,
      "radford et": 27150,
      "al 2019": 1682,
      "toxic content": 33308,
      "content work": 5877,
      "work provides": 35764,
      "test bed": 32760,
      "need better": 22622,
      "data selection": 6854,
      "models recently": 21910,
      "recently neural": 27612,
      "lms demonstrated": 19080,
      "impressive abilities": 14229,
      "generating highquality": 12427,
      "recent papers": 27539,
      "knowledge paper": 15885,
      "paper propose": 24106,
      "method quantitatively": 19963,
      "quantitatively evaluates": 26999,
      "neural lms": 22730,
      "lms understanding": 19119,
      "set linguistic": 29692,
      "linguistic features": 18012,
      "features derived": 11030,
      "transformer lms": 33729,
      "intermediate layer": 15427,
      "layer representations": 17431,
      "gpt2 xlnet": 12970,
      "method shows": 19972,
      "point view": 25066,
      "virtual assistants": 35284,
      "designed allow": 7721,
      "target user": 32060,
      "developed rulebased": 7933,
      "model integrates": 20589,
      "text classification": 32825,
      "classification model": 4600,
      "partofspeech tagging": 24368,
      "approaches including": 2378,
      "trained language": 33404,
      "model gpt": 20553,
      "performed similarly": 24831,
      "faithfulness metrics": 10932,
      "times fewer": 33160,
      "publicly released": 26864,
      "dataset composed": 6953,
      "bidirectional models": 3696,
      "models incremental": 21393,
      "humans process": 13930,
      "process language": 26068,
      "time step": 33145,
      "test models": 32776,
      "models various": 22109,
      "nlu datasets": 22970,
      "datasets compare": 7077,
      "performance using": 24792,
      "evaluation metrics": 9975,
      "metrics results": 20147,
      "results support": 28694,
      "possibility using": 25210,
      "bert model": 3518,
      "model achieves": 20346,
      "achieves better": 1033,
      "training testing": 33631,
      "context available": 5883,
      "contexts generated": 5938,
      "generated language": 12365,
      "model like": 20614,
      "style transfer": 31414,
      "informal formal": 14851,
      "formal language": 11744,
      "word order": 35642,
      "nlp models": 22938,
      "models typically": 22088,
      "work address": 35664,
      "translation problem": 33845,
      "problem build": 25988,
      "new dataset": 22790,
      "dataset parallel": 7019,
      "translation approach": 33820,
      "approach outperforms": 2321,
      "task performed": 32175,
      "computational resource": 5477,
      "promising step": 26298,
      "leveraging machine": 17789,
      "translation models": 33835,
      "transfer code": 33671,
      "code data": 4728,
      "data available": 6618,
      "texttotext transformers": 33020,
      "achieved impressive": 1011,
      "impressive results": 14247,
      "results range": 28666,
      "range natural": 27199,
      "understanding nlu": 34254,
      "generation nlg": 12563,
      "nlg tasks": 22910,
      "tasks current": 32281,
      "pretraining objectives": 25829,
      "objectives masked": 23211,
      "token prediction": 33198,
      "explicitly model": 10549,
      "knowledge everyday": 15848,
      "everyday concepts": 10055,
      "tasks need": 32430,
      "understand generate": 34191,
      "propose generative": 26517,
      "text use": 32960,
      "taskspecific finetuning": 32563,
      "finetuning downstream": 11391,
      "downstream datasets": 8675,
      "datasets furthermore": 7120,
      "furthermore develop": 11993,
      "extensive experimental": 10688,
      "results method": 28643,
      "knowledge parameters": 15886,
      "parameters pretrained": 24278,
      "pretrained texttotext": 25753,
      "texttotext transformer": 33019,
      "relying external": 27976,
      "knowledge graphs": 15859,
      "better performance": 3617,
      "relatively small": 27892,
      "small corpus": 30336,
      "outperforms baseline": 23806,
      "baseline methods": 3253,
      "margin comparable": 19584,
      "comparable larger": 5081,
      "plugandplay method": 25056,
      "method improving": 19932,
      "reasoning ability": 27373,
      "models question": 21883,
      "question answering": 27039,
      "answering recent": 2074,
      "recent works": 27571,
      "works shown": 35826,
      "shown language": 29891,
      "types knowledge": 34065,
      "fail provide": 10906,
      "provide appropriate": 26685,
      "cases paper": 4206,
      "paper ask": 24014,
      "probabilistic models": 25967,
      "t5 bart": 31935,
      "bart gpt2": 3107,
      "qa tasks": 26918,
      "models make": 21697,
      "confidence scores": 5650,
      "outputs inputs": 23891,
      "inputs experiments": 15047,
      "experiments diverse": 10436,
      "diverse range": 8453,
      "range datasets": 27189,
      "datasets demonstrate": 7091,
      "effectiveness methods": 8960,
      "methods perform": 20074,
      "analysis study": 1968,
      "strengths limitations": 31142,
      "shedding light": 29798,
      "calibrating lms": 3975,
      "released code": 27922,
      "gpt2 make": 12914,
      "make models": 19476,
      "models languages": 21419,
      "languages large": 16884,
      "models successful": 22025,
      "english languages": 9484,
      "data computational": 6649,
      "limitations propose": 17932,
      "existing pretrained": 10308,
      "models new": 21750,
      "new languages": 22812,
      "adaptation english": 1179,
      "transformer layers": 33728,
      "layers result": 17445,
      "aligned original": 1742,
      "scale complexity": 29130,
      "embeddings gpt2": 9140,
      "gpt2 small": 12952,
      "small gpt2": 30344,
      "gpt2 medium": 12915,
      "embedding space": 9132,
      "training prevents": 33589,
      "gpt2 english": 12888,
      "sentences generated": 29554,
      "generated gpt2": 12358,
      "model fully": 20540,
      "fully trained": 11958,
      "notoriously difficult": 23048,
      "artificial neural": 2539,
      "generative neural": 12682,
      "recast problem": 27474,
      "language generation": 16082,
      "generation learning": 12539,
      "new paradigm": 22828,
      "network called": 22687,
      "activations pretrained": 1143,
      "model produce": 20728,
      "produce desired": 26143,
      "desired outputs": 7765,
      "original model": 23714,
      "new tasks": 22856,
      "model contribute": 20443,
      "new data": 22789,
      "data set": 6859,
      "loss function": 19244,
      "models control": 21110,
      "autoregressive transformers": 2957,
      "transformers experiments": 33780,
      "experiments stateoftheart": 10484,
      "stateoftheart approaches": 30921,
      "approaches demonstrate": 2372,
      "demonstrate efficacy": 7451,
      "methods using": 20107,
      "using openais": 34865,
      "openais gpt2": 23442,
      "model successfully": 20812,
      "aspects language": 2576,
      "problem using": 26018,
      "datadriven approaches": 6926,
      "existing work": 10324,
      "work does": 35696,
      "increasingly powerful": 14642,
      "models able": 20936,
      "surprisal values": 31762,
      "conducting experiments": 5645,
      "compared existing": 5131,
      "existing baselines": 10263,
      "conditional generation": 5568,
      "sequences models": 29614,
      "knowledge proven": 15894,
      "proven useful": 26677,
      "capture temporal": 4151,
      "temporal relationships": 32699,
      "single model": 30212,
      "sequence use": 29611,
      "generation model": 12550,
      "model capture": 20413,
      "applied different": 2187,
      "different tasks": 8149,
      "model able": 20338,
      "sequences existing": 29613,
      "existing datasets": 10267,
      "evaluation shows": 10011,
      "shows model": 29929,
      "able generate": 745,
      "fit better": 11572,
      "story completion": 31087,
      "completion models": 5261,
      "models pile": 21816,
      "dataset diverse": 6976,
      "diverse text": 8468,
      "text language": 32900,
      "work demonstrated": 35691,
      "dataset diversity": 6977,
      "knowledge downstream": 15838,
      "generalization capability": 12209,
      "english text": 9493,
      "training largescale": 33550,
      "diverse highquality": 8431,
      "existing newly": 10301,
      "performance gpt2": 24618,
      "gpt2 gpt3": 12900,
      "shows models": 29930,
      "trained pile": 33416,
      "improving performance": 14417,
      "performance downstream": 24574,
      "downstream evaluations": 8676,
      "aspects data": 2573,
      "make publicly": 19479,
      "publicly available": 26850,
      "available code": 2968,
      "code used": 4828,
      "success nlp": 31522,
      "model complexity": 20432,
      "computation resources": 5449,
      "extremely long": 10801,
      "long training": 19188,
      "training time": 33633,
      "pretraining finetuning": 25798,
      "finetuning works": 11556,
      "works studied": 35827,
      "model compression": 20434,
      "compression large": 5417,
      "models focusing": 21289,
      "reducing inference": 27751,
      "inference time": 14817,
      "expensive training": 10368,
      "training process": 33592,
      "works use": 35828,
      "extremely large": 10798,
      "large batch": 16930,
      "batch sizes": 3290,
      "pretraining time": 25847,
      "resource demands": 28410,
      "demands paper": 7417,
      "paper inspired": 24058,
      "computer vision": 5505,
      "vision tasks": 35308,
      "tasks propose": 32462,
      "training algorithm": 33439,
      "finetuning largescale": 11437,
      "conduct comprehensive": 5589,
      "finetuning experiments": 11402,
      "tasks results": 32489,
      "achieves comparable": 1036,
      "comparable performance": 5084,
      "performance standard": 24765,
      "code available": 4722,
      "continuous prompts": 5999,
      "generation finetuning": 12506,
      "leverage large": 17752,
      "models perform": 21807,
      "perform downstream": 24483,
      "model parameters": 20689,
      "lightweight alternative": 17837,
      "alternative finetuning": 1851,
      "finetuning natural": 11461,
      "generation tasks": 12613,
      "parameters frozen": 24250,
      "subsequent tokens": 31447,
      "tabletotext generation": 31990,
      "performance data": 24561,
      "outperforms finetuning": 23824,
      "unseen training": 34446,
      "trillion parameter": 33903,
      "models simple": 21980,
      "simple efficient": 30147,
      "mixture experts": 20280,
      "model outrageous": 20678,
      "parameters constant": 24234,
      "constant computational": 5770,
      "computational cost": 5458,
      "despite notable": 7796,
      "widespread adoption": 35591,
      "communication costs": 5050,
      "costs training": 6276,
      "training instability": 33534,
      "switch transformer": 31796,
      "routing algorithm": 28995,
      "improved models": 14316,
      "models reduced": 21914,
      "computational costs": 5461,
      "training techniques": 33629,
      "techniques help": 32641,
      "design models": 7710,
      "models based": 21007,
      "t5base t5large": 31971,
      "t5large obtain": 31976,
      "computational resources": 5478,
      "multilingual settings": 22329,
      "101 languages": 54,
      "languages finally": 16875,
      "advance current": 1417,
      "scale language": 29135,
      "models pretraining": 21849,
      "t5xxl model": 31982,
      "model impact": 20569,
      "multiple parallel": 22407,
      "present indepth": 25535,
      "indepth analysis": 14674,
      "analysis impact": 1927,
      "model user": 20850,
      "input text": 15032,
      "text composition": 32831,
      "compares different": 5186,
      "recent literature": 27530,
      "results reveal": 28672,
      "discuss implications": 8293,
      "research design": 28304,
      "ai instead": 1610,
      "active learning": 1145,
      "classification work": 4620,
      "work propose": 35751,
      "propose use": 26583,
      "learning directly": 17560,
      "easily build": 8794,
      "build machine": 3915,
      "models directly": 21158,
      "data scientists": 6851,
      "approach leverages": 2309,
      "text representation": 32932,
      "like openais": 17889,
      "learning using": 17668,
      "using linear": 34819,
      "linear models": 17990,
      "models providing": 21875,
      "experiments publicly": 10472,
      "datasets empirically": 7101,
      "classification algorithms": 4592,
      "task hand": 32136,
      "targeting specific": 32067,
      "specific issues": 30698,
      "used generate": 34603,
      "prompt learning": 26331,
      "onthefly adaptation": 23374,
      "adaptation unseen": 1195,
      "unseen domains": 34436,
      "address challenging": 1317,
      "domains applied": 8612,
      "labeled unlabeled": 15960,
      "target domain": 32048,
      "domain available": 8555,
      "learning algorithm": 17533,
      "based t5": 3226,
      "t5 language": 31951,
      "model given": 20551,
      "given test": 12775,
      "test example": 32766,
      "trained generate": 33401,
      "token sequence": 33202,
      "domain related": 8590,
      "generated prompt": 12378,
      "semantic space": 29476,
      "domains experiments": 8619,
      "experiments tasks": 10490,
      "tasks text": 32526,
      "sequence tagging": 29608,
      "adaptation scenarios": 1191,
      "substantially outperforms": 31485,
      "outperforms strong": 23857,
      "strong baselines": 31163,
      "nlp systems": 22946,
      "fluent natural": 11635,
      "expert humans": 10510,
      "domain knowledge": 8569,
      "paper make": 24077,
      "main contributions": 19394,
      "present dataset": 25523,
      "new benchmark": 22780,
      "stateoftheart neural": 30965,
      "model achieve": 20343,
      "achieve good": 964,
      "good performance": 12821,
      "second main": 29323,
      "main contribution": 19393,
      "contribution novel": 6037,
      "novel curriculum": 23069,
      "approach model": 2316,
      "related tasks": 27857,
      "models investigate": 21406,
      "investigate model": 15588,
      "t5 exhibits": 31941,
      "consistent human": 5738,
      "considerably improves": 5712,
      "t5 baseline": 31936,
      "bestperforming model": 3584,
      "model fails": 20511,
      "unsolved challenge": 34448,
      "challenge nlp": 4323,
      "systems potential": 31914,
      "gpt2 create": 12879,
      "create synthetic": 6354,
      "synthetic data": 31850,
      "data improve": 6731,
      "prediction performance": 25432,
      "performance nlp": 24692,
      "nlp machine": 22936,
      "learning classification": 17551,
      "classification models": 4601,
      "models use": 22098,
      "input data": 15007,
      "data predict": 6801,
      "predetermined categories": 25402,
      "categories perform": 4225,
      "models require": 21930,
      "require large": 28220,
      "large datasets": 16942,
      "datasets training": 7184,
      "training common": 33450,
      "common practice": 5011,
      "utilize synthetic": 34963,
      "data boost": 6627,
      "using synthetic": 34923,
      "models detect": 21149,
      "created synthetic": 6361,
      "models identify": 21367,
      "learning practitioners": 17630,
      "generate synthetic": 12325,
      "image data": 14064,
      "data train": 6894,
      "convolutional neural": 6123,
      "paper explore": 24045,
      "performance natural": 24687,
      "learning finetune": 17568,
      "finetune pretrained": 11298,
      "transformer model": 33730,
      "reviews data": 28833,
      "data combined": 6644,
      "data create": 6663,
      "create new": 6352,
      "dataset new": 7017,
      "significantly outperformed": 30072,
      "model accuracy": 20342,
      "accuracy precision": 902,
      "extractive abstractive": 10778,
      "language explanations": 16071,
      "method based": 19884,
      "based gpt2": 3167,
      "model perform": 20694,
      "comparative evaluations": 5098,
      "shows promise": 29933,
      "improving pretrained": 14418,
      "social commonsense": 30420,
      "demonstrated outstanding": 7536,
      "outstanding performance": 23903,
      "tasks recently": 32476,
      "reasoning current": 27399,
      "mental states": 19842,
      "improving language": 14411,
      "task requiring": 32189,
      "emotional commonsense": 9204,
      "pretrained roberta": 25743,
      "roberta gpt2": 28917,
      "models propose": 21865,
      "propose architecture": 26497,
      "leveraging external": 17782,
      "optimize model": 23642,
      "achieves competitive": 1042,
      "models provides": 21874,
      "ways improve": 35451,
      "performance particular": 24712,
      "particular task": 24342,
      "models improves": 21375,
      "parallel data": 24169,
      "transfer models": 33681,
      "bart models": 3108,
      "amounts parallel": 1887,
      "task achieve": 32072,
      "achieve new": 973,
      "method unsupervised": 19981,
      "commonsense question": 5037,
      "does rely": 8535,
      "data existing": 6699,
      "popular solution": 25139,
      "solution use": 30482,
      "models score": 21955,
      "score candidate": 29266,
      "candidate choices": 3991,
      "directly conditioned": 8235,
      "question context": 27066,
      "models easily": 21178,
      "mislead model": 20217,
      "present novel": 25544,
      "instead directly": 15120,
      "method generates": 19925,
      "answers generative": 2084,
      "semantic similarity": 29474,
      "verify effectiveness": 35217,
      "effectiveness robustness": 8967,
      "extensive experiments": 10693,
      "experiments evaluate": 10440,
      "evaluate proposed": 9861,
      "proposed method": 26605,
      "datasets method": 7146,
      "method achieves": 19868,
      "synonym replacement": 31810,
      "demonstrates performance": 7562,
      "performance drops": 24578,
      "everyday conversations": 10056,
      "requires understanding": 28265,
      "understanding temporal": 34275,
      "despite recent": 7806,
      "lms t5": 19114,
      "t5 gpt3": 31949,
      "temporal reasoning": 32698,
      "remains largely": 28001,
      "largely underexplored": 17310,
      "underexplored paper": 34142,
      "present study": 25555,
      "study investigate": 31346,
      "reasoning capabilities": 27382,
      "introducing new": 15552,
      "new task": 22855,
      "challenge set": 4328,
      "set timedial": 29710,
      "carefully curated": 4176,
      "results demonstrate": 28586,
      "best performing": 3570,
      "performing models": 24835,
      "task compared": 32094,
      "compared humans": 5141,
      "absolute points": 765,
      "accuracy furthermore": 882,
      "reveals models": 28817,
      "models fail": 21259,
      "dialog context": 8004,
      "based existing": 3155,
      "modeling temporal": 20911,
      "text robust": 32935,
      "contextual reasoning": 5952,
      "reasoning dataset": 27402,
      "dataset publicly": 7027,
      "transformerbased models": 33760,
      "inference speed": 14808,
      "large model": 17231,
      "autoregressive decoding": 2937,
      "decoding process": 7280,
      "accuracy loss": 897,
      "proposed optimization": 26616,
      "optimization techniques": 23636,
      "techniques include": 32644,
      "attention cache": 2714,
      "efficient algorithm": 9023,
      "generation pipeline": 12571,
      "pipeline parallel": 24970,
      "t5 gpt2": 31948,
      "benchmark results": 3409,
      "diverse models": 8439,
      "models demonstrate": 21131,
      "easy use": 8802,
      "use simple": 34568,
      "simple oneline": 30160,
      "source code": 30548,
      "models recent": 21901,
      "size pretrained": 30274,
      "largescale plms": 17372,
      "realworld scenarios": 27347,
      "scenarios present": 29216,
      "techniques use": 32669,
      "finetuning inference": 11421,
      "introduce knowledge": 15512,
      "pretraining process": 25832,
      "existing plms": 10307,
      "training models": 33568,
      "models scratch": 21956,
      "explore best": 10575,
      "best practice": 3574,
      "prompt tuning": 26352,
      "compared conventional": 5128,
      "conventional finetuning": 6073,
      "finetuning prompt": 11496,
      "tuning significantly": 34013,
      "significantly reduces": 30083,
      "reduces number": 27739,
      "number taskspecific": 23162,
      "taskspecific parameters": 32567,
      "implement new": 14159,
      "using largescale": 34816,
      "limited computational": 17944,
      "pretrain models": 25624,
      "models encoderdecoder": 21200,
      "model 11": 20331,
      "11 billion": 68,
      "billion parameters": 3718,
      "parameters experiments": 24243,
      "experiments compare": 10424,
      "tasks experimental": 32321,
      "general language": 12171,
      "inference largescale": 14785,
      "largescale models": 17364,
      "models having": 21348,
      "tens billions": 32716,
      "billions parameters": 3728,
      "parameters single": 24290,
      "single gpu": 30203,
      "largescale knowledge": 17357,
      "enhanced pretraining": 9537,
      "pretraining language": 25805,
      "understanding generation": 34227,
      "generation pretrained": 12574,
      "results various": 28705,
      "gpt3 shown": 13006,
      "scaling pretrained": 29180,
      "models improve": 21372,
      "improve generalization": 14268,
      "generalization abilities": 12204,
      "175 billion": 172,
      "parameters shows": 24287,
      "shows strong": 29938,
      "learning capabilities": 17547,
      "capabilities despite": 4010,
      "despite success": 7817,
      "plain texts": 24991,
      "introducing knowledge": 15551,
      "linguistic knowledge": 18014,
      "trained autoregressive": 33383,
      "traditional finetuning": 33346,
      "finetuning approach": 11372,
      "approach demonstrates": 2256,
      "performance solving": 24759,
      "solving downstream": 30508,
      "tasks order": 32437,
      "solve problems": 30495,
      "problems propose": 26032,
      "propose unified": 26580,
      "unified framework": 34328,
      "framework named": 11884,
      "pretraining largescale": 25814,
      "network trained": 22703,
      "trained model": 33412,
      "model easily": 20477,
      "tasks zeroshot": 32555,
      "learning fewshot": 17566,
      "fewshot learning": 11110,
      "learning finetuning": 17569,
      "finetuning trained": 11550,
      "10 billion": 26,
      "knowledge graph": 15857,
      "results model": 28645,
      "outperforms stateoftheart": 23852,
      "stateoftheart models": 30956,
      "tasks english": 32311,
      "july 2021": 15731,
      "surpassing human": 31756,
      "performance 08": 24512,
      "spanish language": 30594,
      "work presents": 35750,
      "resources available": 28429,
      "industry research": 14746,
      "research community": 28299,
      "models spanish": 21994,
      "pretrained using": 25774,
      "assessed performance": 2606,
      "performance models": 24681,
      "models existing": 21230,
      "existing evaluation": 10271,
      "evaluation datasets": 9937,
      "extractive question": 10779,
      "answering dataset": 2059,
      "dataset created": 6966,
      "outperform existing": 23773,
      "models variety": 22108,
      "nlu tasks": 22975,
      "tasks training": 32535,
      "training settings": 33613,
      "dataset model": 7012,
      "parallel corpus": 24168,
      "based large": 3185,
      "model t5": 20821,
      "t5 trained": 31965,
      "shown produce": 29905,
      "widely applied": 35571,
      "target task": 32058,
      "able improve": 746,
      "performance pretrained": 24718,
      "models unclear": 22091,
      "works previous": 35822,
      "intermediate tasks": 15434,
      "tasks involving": 32382,
      "reasoning work": 27465,
      "reasoning complex": 27394,
      "skills simple": 30315,
      "target tasks": 32059,
      "tasks conduct": 32272,
      "experiments study": 10485,
      "impact different": 14121,
      "findings suggest": 11256,
      "news headlines": 22883,
      "detecting implicit": 7857,
      "causal relations": 4246,
      "task requires": 32188,
      "knowledge existing": 15849,
      "datasets focused": 7119,
      "causal reasoning": 4245,
      "dataset detecting": 6974,
      "dataset includes": 7001,
      "pairs english": 23978,
      "english news": 9487,
      "set models": 29696,
      "models experiments": 21237,
      "including multilingual": 14505,
      "based model": 3196,
      "gpt2 based": 12873,
      "effects prediction": 8983,
      "semantic expansion": 29455,
      "intent detection": 15373,
      "detection based": 7862,
      "challenging task": 4398,
      "additional information": 1256,
      "information regarding": 14902,
      "enhance performance": 9524,
      "detection propose": 7877,
      "generation ability": 12450,
      "context prompt": 5908,
      "apply zeroshot": 2215,
      "zeroshot oneshot": 35986,
      "learning lastly": 17596,
      "queries finetune": 27021,
      "bert roberta": 3527,
      "results performance": 28654,
      "models complex": 21075,
      "complex tasks": 5297,
      "paper demonstrates": 24031,
      "small number": 30361,
      "number examples": 23141,
      "examples specifically": 10143,
      "specifically finetune": 30739,
      "finetune gptneo": 11284,
      "examples finetuning": 10124,
      "gptneo model": 13158,
      "80 accuracy": 553,
      "accuracy achieved": 862,
      "dataset finetuning": 6993,
      "finetuning changes": 11378,
      "algorithm results": 1713,
      "results suggest": 28692,
      "suggest finetuning": 31569,
      "models small": 21988,
      "training machine": 33558,
      "coax models": 4718,
      "complex multistep": 5281,
      "multistep tasks": 22443,
      "tasks investigating": 32380,
      "learning ability": 17531,
      "transfer model": 33680,
      "transformerbased pretrained": 33765,
      "conventional nlp": 6076,
      "tasks struggle": 32515,
      "struggle tasks": 31244,
      "numerical understanding": 23180,
      "possible reasons": 25216,
      "specifically designed": 30735,
      "investigate ability": 15571,
      "t5 models": 31956,
      "tasks improving": 32361,
      "task models": 32161,
      "intermediate training": 15435,
      "training strategy": 33625,
      "strategy enhance": 31119,
      "models performance": 21811,
      "performance text": 24783,
      "specific domains": 30690,
      "strategy includes": 31125,
      "includes novel": 14453,
      "selfsupervised training": 29441,
      "training objective": 33577,
      "improve models": 14279,
      "preliminary experiments": 25488,
      "experiments shown": 10481,
      "approach able": 2225,
      "able outperform": 748,
      "outperform baselines": 23765,
      "domains language": 8622,
      "lms exhibit": 19084,
      "sentence completions": 29530,
      "methods targeted": 20100,
      "targeted syntactic": 32065,
      "makes possible": 19494,
      "possible explore": 25214,
      "apply method": 2208,
      "study behavior": 31305,
      "lms gpt2": 19089,
      "human sentence": 13862,
      "sentence processing": 29541,
      "areas improvement": 2486,
      "improvement truthfulqa": 14349,
      "mimic human": 20170,
      "propose benchmark": 26499,
      "generating answers": 12409,
      "answers questions": 2086,
      "benchmark comprises": 3361,
      "categories including": 4224,
      "law finance": 17419,
      "avoid generating": 3037,
      "imitating human": 14108,
      "t5based model": 31973,
      "model best": 20396,
      "models generated": 21311,
      "models generally": 21306,
      "tasks performance": 32449,
      "performance improves": 24629,
      "improves model": 14384,
      "learned training": 17524,
      "training distribution": 33500,
      "scaling models": 29177,
      "models promising": 21862,
      "finetuning using": 11552,
      "using training": 34930,
      "training objectives": 33578,
      "named entity": 22485,
      "entity recognition": 9648,
      "recognition ner": 27638,
      "seen significant": 29365,
      "significant progress": 30010,
      "progress recent": 26226,
      "stateoftheart sota": 30992,
      "sota models": 30535,
      "models achieving": 20954,
      "achieving high": 1092,
      "high performance": 13576,
      "studies focused": 31269,
      "context paper": 5905,
      "paper introduce": 24059,
      "task aims": 32076,
      "aims generate": 1666,
      "generate relevant": 12317,
      "relevant context": 27937,
      "entities sentence": 9640,
      "sentence context": 29531,
      "facilitate research": 10843,
      "research task": 28365,
      "task present": 32179,
      "dataset consists": 6960,
      "publicly traded": 26866,
      "traded companies": 33333,
      "dataset largest": 7008,
      "35 tokens": 342,
      "tokens sentence": 33244,
      "sentence making": 29538,
      "making challenging": 19499,
      "challenging dataset": 4380,
      "dataset propose": 7025,
      "propose baseline": 26498,
      "baseline approach": 3241,
      "approach combines": 2250,
      "generation algorithm": 12456,
      "model achieving": 20353,
      "dataset evaluate": 6981,
      "evaluate models": 9851,
      "supervised finetuning": 31674,
      "achieve sota": 987,
      "sota results": 30539,
      "results downstream": 28600,
      "model surpasses": 20817,
      "15 points": 143,
      "dataset generated": 6996,
      "encourage research": 9398,
      "research direction": 28308,
      "leading development": 17474,
      "development sophisticated": 7976,
      "models financial": 21272,
      "financial text": 11223,
      "text analysis": 32816,
      "exposure bias": 10636,
      "bias work": 3667,
      "work focuses": 35713,
      "bias text": 3665,
      "despite long": 7795,
      "numerous studies": 23188,
      "impact text": 14138,
      "widelyused pretrained": 35580,
      "gpt2 recently": 12946,
      "holtzman et": 13747,
      "al 2020": 1686,
      "paper attempt": 24015,
      "inspecting hidden": 15087,
      "hidden states": 13539,
      "states gpt2": 31011,
      "results text": 28697,
      "bias study": 3663,
      "study provides": 31385,
      "annotated data": 2017,
      "existing text": 10319,
      "mainly focus": 19403,
      "set realworld": 29704,
      "realworld applications": 27334,
      "applications require": 2176,
      "extending new": 10667,
      "number samples": 23160,
      "requirements introduce": 28239,
      "new problem": 22833,
      "data instead": 6738,
      "finegrained human": 11273,
      "human annotations": 13791,
      "human guidance": 13824,
      "pretrained generative": 25648,
      "specifically propose": 30754,
      "task furthermore": 32128,
      "furthermore devise": 11995,
      "problem setting": 26013,
      "framework uses": 11902,
      "data training": 6896,
      "training classifier": 33448,
      "unlabeled data": 34381,
      "data model": 6770,
      "model refinement": 20748,
      "experiments case": 10423,
      "case studies": 4195,
      "realworld datasets": 27338,
      "demonstrate superior": 7503,
      "superior performance": 31649,
      "performance sota": 24760,
      "zeroshot classification": 35962,
      "lot attention": 19260,
      "nlp domain": 22931,
      "stateoftheart performance": 30969,
      "tasks success": 32518,
      "data large": 6749,
      "large number": 17252,
      "number parameters": 23155,
      "parameters despite": 24239,
      "despite superior": 7819,
      "especially fewshot": 9734,
      "fewshot zeroshot": 11126,
      "compression techniques": 5429,
      "gpt models": 12857,
      "literature work": 18046,
      "work use": 35797,
      "small portion": 30364,
      "finetuned downstream": 11307,
      "tasks using": 32545,
      "evaluate model": 9850,
      "understanding evaluation": 34220,
      "evaluation benchmark": 9924,
      "benchmark tasks": 3416,
      "efficient pretraining": 9054,
      "outperforms existing": 23819,
      "significantly short": 30085,
      "wide range": 35549,
      "large used": 17285,
      "edge devices": 8820,
      "topic model": 33284,
      "increasing attention": 14618,
      "community existing": 5059,
      "existing works": 10327,
      "works focus": 35813,
      "best knowledge": 3559,
      "decoderbased models": 7250,
      "investigated paper": 15603,
      "paper aims": 24010,
      "specifically explore": 30738,
      "current stateoftheart": 6532,
      "stateoftheart knowledge": 30935,
      "techniques improve": 32642,
      "compressed model": 5406,
      "achieve better": 947,
      "performance finetuned": 24601,
      "tasks demonstrate": 32286,
      "impact data": 14120,
      "data cleaning": 6632,
      "model performance": 20695,
      "semantic parsing": 29462,
      "tuning recently": 34011,
      "recently emerged": 27593,
      "emerged effective": 9157,
      "effective method": 8883,
      "adapting pretrained": 1215,
      "models number": 21761,
      "number language": 23148,
      "tuning semantic": 34012,
      "language utterances": 16853,
      "meaning representations": 19723,
      "significantly outperforms": 30074,
      "conduct ablation": 5582,
      "ablation studies": 735,
      "studies different": 31266,
      "different model": 8104,
      "model scales": 20766,
      "increasing model": 14622,
      "model scale": 20764,
      "tuned t5": 33963,
      "improves language": 14379,
      "model generalization": 20542,
      "capabilities led": 4036,
      "like gpt3": 17869,
      "gpt3 t5": 13007,
      "t5 research": 31962,
      "research large": 28330,
      "new model": 22823,
      "training tasks": 33628,
      "engineering efforts": 9465,
      "scale model": 29141,
      "model capacity": 20412,
      "dataset size": 7038,
      "little work": 18054,
      "work improve": 35721,
      "models better": 21023,
      "recently proposed": 27614,
      "substantially improve": 31482,
      "generalization language": 12218,
      "models computational": 21082,
      "computational overhead": 5473,
      "natural questions": 22580,
      "particularly large": 24349,
      "large gains": 16948,
      "gains training": 12073,
      "tasks limited": 32404,
      "sequence sequence": 29606,
      "sequence model": 29603,
      "model extracting": 20510,
      "systems need": 31908,
      "gpt2 shown": 12950,
      "shown promising": 29907,
      "promising results": 26296,
      "nlp benchmarks": 22923,
      "benchmarks like": 3454,
      "research gap": 28320,
      "gap propose": 12105,
      "network architectures": 22685,
      "evaluation best": 9928,
      "performance improvement": 24625,
      "improvement 12": 14322,
      "test set": 32784,
      "set compared": 29677,
      "outperforms bert": 23810,
      "efficient tuning": 9064,
      "tuning pretrained": 34007,
      "models central": 21050,
      "starting point": 30900,
      "range downstream": 27192,
      "models grow": 21342,
      "175b parameters": 176,
      "parameters gpt3": 24253,
      "gpt3 finetuning": 12993,
      "finetuning process": 11495,
      "finetuned model": 11342,
      "finetuned models": 11345,
      "models deployed": 21140,
      "deployed resourceconstrained": 7636,
      "resourceconstrained environments": 28419,
      "propose framework": 26515,
      "parameterefficient finetuning": 24206,
      "finetuning leveraging": 11438,
      "weight updates": 35498,
      "final model": 11178,
      "model weights": 20865,
      "proposed framework": 26599,
      "framework dubbed": 11840,
      "parameter efficient": 24179,
      "efficient finetuning": 9034,
      "lowrank updates": 19309,
      "pretrained weights": 25780,
      "resourceefficient inference": 28424,
      "model leverage": 20611,
      "sparse patterns": 30619,
      "unified approach": 34324,
      "approach extensive": 2279,
      "backbones bert": 3058,
      "gpt2 dozens": 12884,
      "dozens datasets": 8705,
      "datasets consistently": 7083,
      "consistently demonstrate": 5747,
      "demonstrate impressive": 7462,
      "maintaining competitive": 19421,
      "downstream performance": 8684,
      "performance instance": 24631,
      "achieving comparable": 1088,
      "trainable parameters": 33380,
      "parameters bert": 24228,
      "codes available": 4849,
      "model finetuning": 20534,
      "finetuning language": 11424,
      "transformers architecture": 33776,
      "making language": 19506,
      "modeling effective": 20891,
      "nlp task": 22947,
      "leading significant": 17484,
      "significant advancements": 29954,
      "advancements field": 1462,
      "respect input": 28451,
      "input length": 15016,
      "length presents": 17709,
      "presents challenge": 25574,
      "long texts": 19187,
      "propose finetuning": 26514,
      "finetuning framework": 11407,
      "architecture current": 2439,
      "current pretrained": 6524,
      "models incorporate": 21385,
      "entity representations": 9653,
      "available information": 2982,
      "information outside": 14894,
      "model results": 20757,
      "better language": 3610,
      "fraction computational": 11816,
      "implement approach": 14157,
      "approach using": 2356,
      "gpt2 compare": 12878,
      "model original": 20667,
      "proposed model": 26612,
      "lower perplexity": 19289,
      "finetuned version": 11362,
      "compare models": 5109,
      "performance terms": 24781,
      "terms accuracy": 32737,
      "coreference annotations": 6156,
      "fewshot text": 11124,
      "augmentation techniques": 2806,
      "enhancing performance": 9573,
      "data sparsity": 6877,
      "stateoftheart generative": 30932,
      "shown provide": 29910,
      "provide significant": 26729,
      "augmentation text": 2807,
      "classification tasks": 4614,
      "tasks fewshot": 32329,
      "fewshot settings": 11123,
      "fully explored": 11954,
      "domains paper": 8632,
      "paper leverage": 24075,
      "generating artificial": 12411,
      "order improve": 23675,
      "improve classification": 14257,
      "classification performance": 4602,
      "performance aim": 24522,
      "selection process": 29394,
      "classifier performance": 4625,
      "perform experiments": 24486,
      "seed selection": 29354,
      "selection strategies": 29395,
      "domain expert": 8561,
      "results finetuning": 28617,
      "outperform competitive": 23768,
      "competitive baselines": 5221,
      "baselines finally": 3265,
      "interesting research": 15410,
      "research avenues": 28296,
      "scalable efficient": 29125,
      "optimization method": 23631,
      "residual learning": 28391,
      "learning scheme": 17651,
      "dynamically adjust": 8767,
      "test time": 32792,
      "enhancement performance": 9541,
      "incurring minimal": 14662,
      "memory training": 19832,
      "training overhead": 33581,
      "demonstrate proposed": 7481,
      "performance degradation": 24563,
      "trained endtoend": 33395,
      "generation using": 12631,
      "current language": 6500,
      "models generate": 21307,
      "generate highquality": 12285,
      "highquality text": 13701,
      "tease apart": 32602,
      "sequential structure": 29627,
      "models lstm": 21692,
      "lstm transformer": 19328,
      "transformerxl gpt2": 33802,
      "text substantially": 32953,
      "text models": 32912,
      "models test": 22054,
      "1000 words": 45,
      "words long": 35658,
      "set perform": 29702,
      "perform extensive": 24487,
      "extensive manual": 10710,
      "manual analysis": 19553,
      "analysis showing": 1965,
      "novel text": 23119,
      "efficient sparse": 9060,
      "sparse training": 30620,
      "training neural": 33571,
      "network models": 22699,
      "networks generalize": 22711,
      "expensive train": 10367,
      "ideally like": 13988,
      "reduce computational": 27702,
      "sparse model": 30618,
      "model training": 20839,
      "promising approach": 26282,
      "approach achieve": 2226,
      "challenges existing": 4342,
      "existing methods": 10293,
      "methods struggle": 20098,
      "model components": 20433,
      "modern hardware": 22158,
      "uses simple": 34715,
      "lowrank matrices": 19306,
      "network layers": 22696,
      "empirically validate": 9251,
      "speeds training": 30799,
      "accuracyefficiency tradeoffs": 918,
      "models train": 22064,
      "25x faster": 277,
      "faster dense": 10998,
      "vision transformer": 35309,
      "drop accuracy": 8740,
      "conditional language": 5569,
      "models catastrophic": 21048,
      "learning shifting": 17655,
      "generalpurpose pretrained": 12257,
      "trained selfsupervised": 33424,
      "amounts data": 1884,
      "applied solve": 2198,
      "number tasks": 23161,
      "training methodology": 33564,
      "abstractive summarization": 777,
      "code generation": 4754,
      "raises important": 27167,
      "important question": 14208,
      "adapt pretrained": 1168,
      "models meet": 21713,
      "meet requirements": 19780,
      "general capabilities": 12158,
      "work proposed": 35761,
      "solve problem": 30494,
      "policy gradients": 25083,
      "despite effectiveness": 7774,
      "effectiveness approach": 8939,
      "paper extend": 24051,
      "different control": 8061,
      "translation summarization": 33850,
      "baseline approaches": 3242,
      "does result": 8540,
      "result catastrophic": 28542,
      "systems use": 31921,
      "use large": 34542,
      "large neural": 17246,
      "networks require": 22714,
      "resources training": 28448,
      "inspired human": 15095,
      "knowledge acquisition": 15810,
      "researchers proposed": 28380,
      "curriculum learning": 6548,
      "facilitate training": 10847,
      "training work": 33644,
      "work investigates": 35729,
      "learning large": 17592,
      "modern language": 22159,
      "bert t5": 3533,
      "based range": 3215,
      "range complexity": 27188,
      "sampling strategies": 29096,
      "experiments different": 10435,
      "based various": 3237,
      "random sampling": 27177,
      "performs better": 24842,
      "crosslingual transfer": 6420,
      "building block": 3922,
      "nlp applications": 22921,
      "models requires": 21931,
      "existing models": 10299,
      "trained english": 33396,
      "alleviate problem": 1790,
      "problem introduce": 25993,
      "introduce novel": 15524,
      "efficiently effectively": 9070,
      "effectively transfer": 8932,
      "model uses": 20851,
      "subwordbased tokenization": 31504,
      "learns embedding": 17678,
      "source model": 30567,
      "model english": 20489,
      "target language": 32051,
      "token embeddings": 33190,
      "semantically similar": 29486,
      "static word": 31013,
      "method lowresource": 19944,
      "lowresource languages": 19313,
      "proposed methods": 26611,
      "outperforms models": 23834,
      "models comparable": 21069,
      "comparable size": 5090,
      "method makes": 19945,
      "make code": 19454,
      "code models": 4786,
      "models publicly": 21877,
      "available efficient": 2973,
      "adaptation pretrained": 1187,
      "models remarkable": 21927,
      "remarkable success": 28056,
      "success large": 31514,
      "models driven": 21176,
      "dense models": 7607,
      "trained massive": 33410,
      "unlabeled unstructured": 34384,
      "source text": 30570,
      "transferring knowledge": 33691,
      "domain typically": 8601,
      "introduce method": 15515,
      "adaptation diverse": 1177,
      "diverse domains": 8425,
      "domains using": 8643,
      "computationally efficient": 5485,
      "efficient adapter": 9022,
      "adapter approach": 1199,
      "based observation": 3202,
      "frozen pretrained": 11939,
      "model approach": 20373,
      "approach enables": 2269,
      "results gpt2": 28620,
      "gpt2 large": 12912,
      "large fraction": 16947,
      "additionally provide": 1298,
      "time algorithm": 33108,
      "cost inference": 6245,
      "inference latency": 14786,
      "transformer encoder": 33715,
      "efficient architecture": 9026,
      "paper proposes": 24120,
      "proposes efficient": 26627,
      "efficient transformer": 9063,
      "inference computational": 14766,
      "desired inference": 7764,
      "finetuning phase": 11483,
      "encoder layer": 9349,
      "using proposed": 34886,
      "proposed attention": 26594,
      "property inference": 26480,
      "range inference": 27195,
      "inference speedup": 14809,
      "training proposed": 33595,
      "method applied": 19878,
      "bertbase gpt2": 3539,
      "models evaluation": 21221,
      "evaluation extensive": 9946,
      "higher transformer": 13611,
      "latency experimental": 17403,
      "results extensive": 28611,
      "sentiment analysis": 29566,
      "classification text": 4618,
      "like glue": 17863,
      "method effective": 19907,
      "various datasets": 35085,
      "minimal impact": 20188,
      "accuracy drop": 871,
      "models llms": 21472,
      "necessary training": 22610,
      "foundation models": 11798,
      "automatic code": 2876,
      "code summarization": 4819,
      "software development": 30454,
      "help reduce": 13512,
      "manual writing": 19560,
      "artificial intelligence": 2533,
      "paradigm shift": 24159,
      "pretrained massive": 25710,
      "massive data": 19625,
      "data finetuned": 6713,
      "instead learning": 15123,
      "robust approach": 28931,
      "approach automatic": 2239,
      "summarization based": 31609,
      "based neural": 3201,
      "single neural": 30216,
      "neural model": 22736,
      "model named": 20651,
      "gaussian noise": 12133,
      "contextual information": 5950,
      "information optimize": 14893,
      "furthermore introduce": 12004,
      "knowledge transfer": 15915,
      "continuous pretraining": 5998,
      "tasks general": 32345,
      "finally evaluate": 11193,
      "dataset code": 6949,
      "sequencetosequence model": 29619,
      "model simple": 20786,
      "generation recent": 12590,
      "approaches proposed": 2387,
      "dedicated training": 7305,
      "training paradigms": 33584,
      "decoding strategies": 7282,
      "commonly used": 5029,
      "seq2seq language": 29589,
      "model bart": 20385,
      "easily adapted": 8792,
      "single batch": 30200,
      "using simple": 34910,
      "simple training": 30165,
      "training procedure": 33590,
      "results benchmarks": 28575,
      "benchmarks approach": 3431,
      "existing stateoftheart": 10315,
      "benchmarks test": 3475,
      "test abilities": 32757,
      "models difficult": 21157,
      "models exploit": 21239,
      "adversarial examples": 1514,
      "make errors": 19468,
      "data construction": 6657,
      "leads enhanced": 17492,
      "user engagement": 34646,
      "collected data": 4921,
      "collect highquality": 4917,
      "highquality data": 13684,
      "data scale": 6846,
      "using method": 34839,
      "method create": 19895,
      "yesno questions": 35904,
      "questions demonstrate": 27104,
      "best baseline": 3555,
      "achieves accuracy": 1028,
      "substantially higher": 31481,
      "knowledge grounding": 15862,
      "texttotext language": 33012,
      "models structured": 22015,
      "complete user": 5252,
      "answering knowledge": 2063,
      "knowledge bases": 15817,
      "inputs outputs": 15050,
      "studied separately": 31260,
      "paper overcome": 24083,
      "overcome limitation": 23922,
      "framework unifies": 11900,
      "tasks texttotext": 32530,
      "texttotext format": 33011,
      "single task": 30224,
      "task domain": 32112,
      "t5 different": 31940,
      "different sizes": 8139,
      "simple modifications": 30158,
      "achieves stateoftheart": 1070,
      "improves performance": 14387,
      "performance tasks": 24780,
      "tasks largely": 32394,
      "overall performance": 23910,
      "zeroshot fewshot": 35971,
      "gpt3 codex": 12988,
      "conduct series": 5618,
      "controlled experiments": 6063,
      "knowledge encoding": 15842,
      "model fairness": 20512,
      "explored paper": 10609,
      "paper examine": 24042,
      "examine effect": 10099,
      "distillation pruning": 8347,
      "pruning toxicity": 26816,
      "toxicity bias": 33317,
      "pruning methods": 26814,
      "methods gpt2": 20043,
      "model consistent": 20438,
      "line research": 17982,
      "serves reference": 29655,
      "compressed models": 5407,
      "models extends": 21246,
      "new ways": 22866,
      "written language": 35863,
      "ai technologies": 1619,
      "technologies like": 32678,
      "new tools": 22860,
      "new concept": 22787,
      "ai technology": 1620,
      "generate humanlike": 12288,
      "humanlike text": 13909,
      "aigenerated content": 1633,
      "article introduces": 2522,
      "models end": 21203,
      "end paper": 9414,
      "paper focuses": 24053,
      "processing models": 26113,
      "need access": 22617,
      "access training": 827,
      "parameters training": 24293,
      "essential enhancing": 9758,
      "learning ml": 17605,
      "performance recent": 24734,
      "recent empirical": 27518,
      "empirical studies": 9239,
      "conduct largescale": 5614,
      "analysis neural": 1944,
      "type model": 34054,
      "model selection": 20771,
      "test performance": 32778,
      "performance paper": 24709,
      "tasks prior": 32456,
      "prior work": 25941,
      "vision cv": 35292,
      "tasks ii": 32356,
      "directly predict": 8239,
      "data compute": 6650,
      "able provide": 752,
      "provide model": 26714,
      "results large": 28637,
      "transformers trained": 33798,
      "trained different": 33389,
      "different settings": 8138,
      "metrics derived": 20134,
      "particularly useful": 24357,
      "extend prior": 10657,
      "power law": 25323,
      "largescale training": 17382,
      "techniques enable": 32637,
      "remain open": 27985,
      "open question": 23416,
      "pretraining bert": 25786,
      "gpt paper": 12862,
      "paper demonstrate": 24030,
      "slow convergence": 30332,
      "applied alleviate": 2186,
      "limitation propose": 17913,
      "optimizer states": 23651,
      "wallclock time": 35413,
      "gpt2 pretraining": 12941,
      "able reduce": 753,
      "higher training": 13609,
      "training throughput": 33632,
      "endtoend training": 9443,
      "compared stateoftheart": 5172,
      "stateoftheart baseline": 30923,
      "convergence speed": 6085,
      "accuracy glue": 885,
      "validation set": 35007,
      "alternative approach": 1849,
      "propose alternative": 26494,
      "network training": 22704,
      "training using": 33642,
      "seminal work": 29497,
      "models glm": 21319,
      "problem solving": 26015,
      "approach leads": 2308,
      "special cases": 30655,
      "layer pretrained": 17429,
      "approach used": 2351,
      "essential step": 9761,
      "models llm": 21469,
      "llm demonstrate": 18285,
      "demonstrate applicability": 7435,
      "graph neural": 13225,
      "performance approach": 24524,
      "approach compared": 2251,
      "gradient descent": 13187,
      "synthetic real": 31858,
      "prediction tasks": 25438,
      "performance metrics": 24677,
      "mixtureofexperts moe": 20287,
      "moe models": 22198,
      "models allow": 20974,
      "parameters greatly": 24255,
      "given sample": 12769,
      "fixed number": 11578,
      "token using": 33205,
      "relative importance": 27879,
      "address propose": 1353,
      "topk experts": 33294,
      "systematically study": 31883,
      "using computational": 34754,
      "method improves": 19930,
      "improves training": 14396,
      "2x computational": 304,
      "cost method": 6249,
      "method demonstrates": 19898,
      "higher performance": 13600,
      "performance finetuning": 24603,
      "selected tasks": 29383,
      "glue superglue": 12803,
      "method outperforms": 19951,
      "outperforms t5": 23859,
      "dense model": 7606,
      "11 tasks": 74,
      "models building": 21035,
      "highly capable": 13658,
      "capable language": 4109,
      "years despite": 35890,
      "despite great": 7780,
      "great performance": 13252,
      "incur high": 14658,
      "high computational": 13558,
      "model desirable": 20463,
      "performance case": 24538,
      "compression paper": 5421,
      "dynamic inference": 8760,
      "inference approach": 14761,
      "approach called": 2246,
      "inference large": 14781,
      "decision making": 7233,
      "latent space": 17410,
      "method easily": 19906,
      "unlike existing": 34394,
      "tasks method": 32418,
      "method works": 19985,
      "sequencetosequence tasks": 29622,
      "set experiments": 29686,
      "experiments t5": 10488,
      "t5 bert": 31937,
      "code demo": 4746,
      "demo available": 7420,
      "paradigm finetuning": 24156,
      "models parameterefficient": 21801,
      "explore use": 10603,
      "learn taskspecific": 17514,
      "feature maps": 11025,
      "time enabling": 33121,
      "enabling flexible": 9321,
      "information sharing": 14913,
      "multitask learning": 22450,
      "parameters achieving": 24224,
      "computational efficiency": 5466,
      "extensive empirical": 10682,
      "empirical experiments": 9225,
      "experiments demonstrate": 10431,
      "achieve superior": 995,
      "superior performances": 31655,
      "understanding benchmarks": 34212,
      "architecture pretrained": 2451,
      "moe architecture": 22196,
      "achieved remarkable": 1013,
      "parameters base": 24226,
      "model extended": 20505,
      "layer increase": 17425,
      "increase model": 14598,
      "sharing parameters": 29791,
      "core information": 6154,
      "information different": 14860,
      "experiments based": 10421,
      "gpt2 improved": 12907,
      "improved performance": 14317,
      "performance efficiency": 24580,
      "total parameters": 33299,
      "superior model": 31648,
      "performance compared": 24550,
      "code publicly": 4801,
      "unclear extent": 34124,
      "training corpus": 33459,
      "paraphrased sentences": 24303,
      "similar training": 30120,
      "training samples": 33604,
      "work study": 35790,
      "gpt2 generated": 12894,
      "generated texts": 12394,
      "texts comparison": 32989,
      "finetuned lms": 11341,
      "domainspecific corpora": 8647,
      "extensively used": 10719,
      "practice results": 25378,
      "decoding methods": 7277,
      "vary based": 35165,
      "words phrases": 35660,
      "training sets": 33612,
      "ethical implications": 9806,
      "data increase": 6735,
      "raising concerns": 27172,
      "larger models": 17330,
      "larger training": 17339,
      "training corpora": 33458,
      "sensitive information": 29516,
      "data source": 6870,
      "various models": 35121,
      "models proposed": 21869,
      "structures neural": 31232,
      "specific language": 30700,
      "model usually": 20853,
      "network rnn": 22702,
      "models transformer": 22082,
      "gpt2 paper": 12934,
      "models novel": 21760,
      "modeling objective": 20901,
      "probability distribution": 25971,
      "given context": 12744,
      "human evaluations": 13810,
      "easily effectively": 8796,
      "effectively applied": 8911,
      "different neural": 8111,
      "models improving": 21376,
      "generation various": 12638,
      "various tasks": 35149,
      "tasks language": 32387,
      "generation dialogue": 12486,
      "prompt completion": 26313,
      "completion language": 5259,
      "recently shown": 27624,
      "generate factual": 12275,
      "zhou et": 36003,
      "al 2021": 1687,
      "combination retrieval": 4952,
      "method applies": 19879,
      "final response": 11182,
      "response using": 28483,
      "dialogue model": 8016,
      "stateoftheart model": 30955,
      "chen et": 4529,
      "terms consistency": 32742,
      "prompt completions": 26314,
      "standard language": 30878,
      "outperforms gpt2": 23825,
      "2019 gpt3": 223,
      "gpt3 brown": 12983,
      "brown et": 3897,
      "larger model": 17328,
      "model code": 20421,
      "models deep": 21126,
      "learning dl": 17561,
      "finetuning large": 11427,
      "large numbers": 17257,
      "performance task": 24779,
      "alzheimers disease": 1863,
      "disease ad": 8304,
      "questions remain": 27128,
      "ability generalize": 671,
      "generalize small": 12233,
      "available research": 2999,
      "parameters directly": 24241,
      "propose novel": 26543,
      "gpt2 pretrained": 12937,
      "pretrained general": 25647,
      "general english": 12165,
      "approaches stateoftheart": 2395,
      "text data": 32841,
      "data widely": 6913,
      "description task": 7684,
      "conversations furthermore": 6107,
      "generates text": 12404,
      "text characteristics": 32824,
      "better understanding": 3634,
      "understanding relationships": 34268,
      "human speech": 13864,
      "speech language": 30782,
      "outofdistribution generalization": 23748,
      "generalization natural": 12220,
      "nlp algorithms": 22919,
      "remains significant": 28012,
      "significant challenge": 29964,
      "challenge paper": 4324,
      "paper addresses": 24004,
      "addresses issue": 1364,
      "labeled data": 15955,
      "data multiple": 6779,
      "multiple source": 22419,
      "target domains": 32050,
      "training innovative": 33533,
      "innovative framework": 14996,
      "framework employs": 11843,
      "hypernetwork generate": 13956,
      "generate task": 12328,
      "tasks sentiment": 32493,
      "sentiment classification": 29569,
      "advanced version": 1445,
      "demonstrating effectiveness": 7580,
      "use cases": 34516,
      "feedforward layers": 11076,
      "layers build": 17437,
      "vocabulary space": 35384,
      "space transformerbased": 30584,
      "transformerbased language": 33747,
      "modern nlp": 22168,
      "work make": 35735,
      "network ffn": 22690,
      "ffn layers": 11128,
      "layers building": 17438,
      "building blocks": 3923,
      "changing distribution": 4414,
      "distribution vocabulary": 8398,
      "leverage findings": 17747,
      "findings controlling": 11232,
      "lm predictions": 19061,
      "reduce toxicity": 27727,
      "computation efficiency": 5446,
      "early exit": 8774,
      "recent neural": 27537,
      "neural networkbased": 22751,
      "scaling size": 29181,
      "size training": 30288,
      "training datasets": 33497,
      "parameters models": 24274,
      "models scaling": 21953,
      "various factors": 35093,
      "factors including": 10871,
      "data ensure": 6695,
      "results work": 28713,
      "models scale": 21951,
      "data evaluation": 6697,
      "hundreds billions": 13943,
      "datasets multiple": 7151,
      "encoderdecoder models": 9370,
      "decoderonly architectures": 7253,
      "open source": 23421,
      "source available": 30547,
      "efficient accurate": 9020,
      "popular approach": 25112,
      "approach reduce": 2332,
      "reduce compute": 27706,
      "compute memory": 5495,
      "weight matrices": 35495,
      "seen widespread": 29368,
      "address issues": 1336,
      "issues propose": 15672,
      "represent commonly": 28132,
      "optimal solution": 23616,
      "unlock new": 34410,
      "ways train": 35452,
      "finetune sparse": 11300,
      "comparable model": 5082,
      "model quality": 20741,
      "technique called": 32618,
      "serve useful": 29650,
      "intermediate representation": 15430,
      "bert pretraining": 3525,
      "optimized implementation": 23646,
      "mlperf 11": 20306,
      "bert finetuning": 3504,
      "comparable accuracy": 5073,
      "lms shown": 19108,
      "knowledge pretraining": 15889,
      "pretraining corpora": 25789,
      "corpora limited": 6167,
      "factually correct": 10895,
      "generation used": 12629,
      "focus modifying": 11652,
      "pretraining task": 25844,
      "finetuning objectives": 11465,
      "require additional": 28212,
      "lms practical": 19101,
      "practical applications": 25361,
      "novel decoding": 23073,
      "decoding algorithm": 7269,
      "based current": 3148,
      "current context": 6487,
      "local memory": 19133,
      "reinforcement learning": 27838,
      "gpt2 bart": 12872,
      "models particularly": 21806,
      "strong performance": 31185,
      "performance fewshot": 24597,
      "fewshot scenarios": 11121,
      "evaluation confirms": 9932,
      "language input": 16096,
      "input context": 15005,
      "compared multiple": 5155,
      "multiple baselines": 22379,
      "alleviates exposure": 1793,
      "generation quality": 12584,
      "generating longer": 12434,
      "longer sequences": 19201,
      "extraction text": 10775,
      "generation paper": 12566,
      "paper introduces": 24062,
      "generation different": 12488,
      "prior studies": 25940,
      "studies work": 31291,
      "datasets design": 7095,
      "effective model": 8887,
      "tokens context": 33220,
      "annotation data": 2025,
      "scenarios model": 29211,
      "model better": 20398,
      "pretrained t5": 25751,
      "model introduce": 20593,
      "20 billion": 208,
      "billion parameter": 3715,
      "openly available": 23474,
      "available public": 2998,
      "permissive license": 24863,
      "knowledge largest": 15876,
      "autoregressive model": 2952,
      "model publicly": 20738,
      "weights time": 35514,
      "work models": 35736,
      "models architecture": 20988,
      "architecture training": 2455,
      "training evaluate": 33511,
      "evaluate performance": 9854,
      "performance range": 24730,
      "performance evaluated": 24584,
      "similarly sized": 30135,
      "models opensource": 21771,
      "training evaluation": 33513,
      "evaluation code": 9929,
      "fewshot learners": 11109,
      "models successfully": 22026,
      "zero fewshot": 35936,
      "learning paradigms": 17627,
      "opens new": 23478,
      "new possibilities": 22831,
      "gptlike models": 13154,
      "models 13": 20920,
      "13 billion": 107,
      "billion 13": 3713,
      "parameters trained": 24292,
      "languages 25": 16862,
      "language families": 16074,
      "families using": 10972,
      "architecture using": 2457,
      "sparse attention": 30610,
      "attention mechanism": 2725,
      "training inference": 33531,
      "resulting models": 28559,
      "performance par": 24711,
      "recently released": 27619,
      "low resource": 19274,
      "resource languages": 28413,
      "architecture design": 2440,
      "data preparation": 6802,
      "versions model": 35237,
      "measure model": 19734,
      "model perplexity": 20711,
      "multilingual tasks": 22332,
      "tasks including": 32364,
      "including classification": 14467,
      "sequence labeling": 29599,
      "models evaluated": 21219,
      "furthermore compared": 11988,
      "stateoftheart multilingual": 30959,
      "multilingual model": 22319,
      "tasks nlp": 32431,
      "models generalize": 21305,
      "unseen tasks": 34444,
      "tasks provided": 32465,
      "task instructions": 32142,
      "address question": 1354,
      "question introduce": 27070,
      "supernaturalinstructions benchmark": 31666,
      "diverse nlp": 8446,
      "task types": 32204,
      "types including": 34063,
      "including limited": 14497,
      "classification extraction": 4597,
      "large diverse": 16944,
      "collection tasks": 4933,
      "tasks enables": 32308,
      "crosstask generalization": 6427,
      "instructions training": 15277,
      "models follow": 21290,
      "follow instructions": 11677,
      "unseen ones": 34443,
      "ones furthermore": 23348,
      "incontext instructions": 14546,
      "plain language": 24990,
      "language task": 16829,
      "task definitions": 32104,
      "kshot examples": 15944,
      "instructionfollowing models": 15236,
      "models instructgpt": 21399,
      "despite order": 7798,
      "order magnitude": 23676,
      "magnitude smaller": 19384,
      "function various": 11963,
      "scaling parameters": 29179,
      "number observed": 23154,
      "tasks number": 32433,
      "facilitate future": 10840,
      "systems user": 31923,
      "underlying user": 34164,
      "user information": 34652,
      "feature modern": 11026,
      "evaluation systems": 10016,
      "questions requires": 27131,
      "requires significant": 28261,
      "significant human": 29984,
      "human effort": 13802,
      "timeconsuming expensive": 33148,
      "expensive paper": 10364,
      "propose conversational": 26501,
      "conversational user": 6104,
      "user simulator": 34672,
      "simulator called": 30192,
      "automatic evaluation": 2880,
      "automatically answering": 2903,
      "including automated": 14458,
      "automated natural": 2870,
      "generation metrics": 12548,
      "responses generated": 28492,
      "answers make": 2085,
      "make steps": 19484,
      "multiturn interactions": 22465,
      "user goal": 34650,
      "currently available": 6541,
      "available datasets": 2972,
      "data acquisition": 6592,
      "model capable": 20410,
      "capable providing": 4118,
      "providing accurate": 26770,
      "discuss capabilities": 8290,
      "capabilities model": 4044,
      "model multiturn": 20650,
      "provide code": 26688,
      "model used": 20849,
      "used research": 34622,
      "research topic": 28366,
      "bert language": 3514,
      "given sentence": 12770,
      "social media": 30423,
      "media platforms": 19760,
      "extensively studied": 10718,
      "pretrained transformerbased": 25767,
      "transformerbased architectures": 33746,
      "gaining popularity": 12068,
      "data scarce": 6849,
      "models present": 21836,
      "bert models": 3520,
      "using masked": 34836,
      "language modelling": 16228,
      "models subsequent": 22020,
      "pos tagging": 25155,
      "based generative": 3164,
      "generative transformer": 12710,
      "capable generating": 4107,
      "codemixed data": 4844,
      "dataset models": 7013,
      "useful new": 34641,
      "new language": 22811,
      "language learners": 16108,
      "solving common": 30506,
      "does exist": 8526,
      "language present": 16772,
      "indian languages": 14683,
      "languages paper": 16903,
      "propose transformerbased": 26579,
      "learning approach": 17536,
      "approach tackle": 2344,
      "existing systems": 10317,
      "using mt5": 34845,
      "mt5 model": 22256,
      "model architecture": 20374,
      "translation language": 33829,
      "berts masked": 3546,
      "modeling mlm": 20900,
      "behavior transformerbased": 3323,
      "methods focus": 20040,
      "focus probing": 11654,
      "models outside": 21787,
      "input features": 15011,
      "transformerbased lms": 33758,
      "provides finegrained": 26752,
      "models internal": 21403,
      "recent method": 27533,
      "token representations": 33201,
      "demonstrate utility": 7512,
      "model behavior": 20391,
      "process release": 26083,
      "opensource tool": 23545,
      "controlled text": 6065,
      "desirable attributes": 7758,
      "attribute classifiers": 2757,
      "bridge gap": 3864,
      "training prompt": 33593,
      "prompt task": 26347,
      "task testing": 32197,
      "introduces trainable": 15548,
      "generation experiments": 12498,
      "experiments 11": 10416,
      "demonstrate strong": 7500,
      "strong performances": 31188,
      "training parameters": 33585,
      "parameters gpt2": 24252,
      "explanations fewshot": 10541,
      "fewshot prompting": 11119,
      "textual reasoning": 33036,
      "prompting large": 26381,
      "model llm": 20620,
      "llm like": 18334,
      "incontext learning": 14548,
      "learning study": 17658,
      "study question": 31388,
      "tasks involve": 32381,
      "reasoning text": 27461,
      "text question": 32926,
      "answering natural": 2065,
      "inference test": 14815,
      "performance llms": 24663,
      "reasoning datasets": 27403,
      "datasets using": 7187,
      "using prompts": 34885,
      "different styles": 8145,
      "opt gpt3": 23593,
      "gpt3 davinci": 12989,
      "accuracy improvements": 891,
      "generated llms": 12373,
      "models predictions": 21832,
      "factually grounded": 10896,
      "grounded input": 13289,
      "accurate predictions": 927,
      "using automatically": 34737,
      "reliability explanations": 27949,
      "vector representations": 35196,
      "conversational systems": 6103,
      "idioms figurative": 14031,
      "responses prompts": 28505,
      "prompts containing": 26409,
      "languages cultures": 16867,
      "great challenge": 13248,
      "challenge natural": 4319,
      "translation mt": 33836,
      "conversational ai": 6094,
      "tasks investigate": 32379,
      "conversation generation": 6090,
      "generation achieve": 12451,
      "achieve stateoftheart": 990,
      "macro f1": 19377,
      "classification task": 4613,
      "task using": 32208,
      "t5 model": 31954,
      "model dialogue": 20469,
      "evaluated using": 9884,
      "using automatic": 34736,
      "automatic metric": 2888,
      "metric perplexity": 20124,
      "evaluation results": 10000,
      "time compared": 33111,
      "similar model": 30109,
      "huggingface hub": 13787,
      "public access": 26830,
      "coreference resolution": 6158,
      "task understanding": 32205,
      "language large": 16105,
      "benefits large": 3490,
      "largely rely": 17309,
      "rely supervised": 27972,
      "prompt engineering": 26322,
      "engineering paper": 9468,
      "pretrained llms": 25706,
      "llms abilities": 18405,
      "abilities limitations": 632,
      "experiments gpt2": 10445,
      "gpt2 gptneo": 12904,
      "valid answers": 34989,
      "results evaluation": 28606,
      "assessing quality": 2614,
      "texttotext models": 33015,
      "diverse tasks": 8466,
      "tasks datasets": 32284,
      "summarization question": 31622,
      "additionally present": 1296,
      "finetuned various": 11361,
      "tasks single": 32507,
      "single training": 30226,
      "initializing model": 14971,
      "multilingual t5": 22330,
      "t5 mt5": 31957,
      "tasks summarization": 32519,
      "better results": 3625,
      "results encoderdecoder": 28601,
      "encoderdecoder architectures": 9364,
      "growing body": 13312,
      "body work": 3806,
      "work recent": 35771,
      "major problems": 19444,
      "lack systematic": 16005,
      "work revisit": 35776,
      "data size": 6868,
      "size model": 30264,
      "model incorporating": 20579,
      "characterlevel information": 4435,
      "release new": 27916,
      "bertstyle models": 3552,
      "comprehensive empirical": 5364,
      "systematically evaluate": 31879,
      "performance existing": 24587,
      "benchmark arabic": 3354,
      "significantly outperform": 30070,
      "plms achieve": 25039,
      "discriminative generative": 8286,
      "models source": 21992,
      "code reproduce": 4811,
      "reproduce results": 28200,
      "results available": 28573,
      "understand language": 34193,
      "works usually": 35830,
      "usually focus": 34947,
      "model families": 20513,
      "architecture pretraining": 2452,
      "datasets diverse": 7098,
      "diverse set": 8459,
      "including t5": 14519,
      "additionally adapt": 1269,
      "models evaluate": 21218,
      "evaluate gpt": 9839,
      "networks different": 22710,
      "findings models": 11243,
      "models resolve": 21935,
      "zeroshot fashion": 35970,
      "using existing": 34775,
      "existing pretraining": 10311,
      "global model": 12797,
      "size dataset": 30244,
      "dataset pretraining": 7022,
      "pretraining objective": 25827,
      "linguistic capabilities": 18007,
      "textual explanations": 33029,
      "inference nli": 14795,
      "current benchmarks": 6485,
      "spurious correlations": 30835,
      "tackle problem": 31999,
      "problem work": 26019,
      "models right": 21946,
      "language making": 16113,
      "address issue": 1331,
      "collect data": 4913,
      "framework based": 11832,
      "crowd workers": 6430,
      "expert annotators": 10507,
      "human annotators": 13793,
      "linguistic phenomena": 18017,
      "baseline performance": 3256,
      "performance t5": 24777,
      "step closer": 31038,
      "developing models": 7945,
      "models understand": 22093,
      "models conditional": 21087,
      "input sequence": 15027,
      "sequence tokens": 29610,
      "set nlp": 29699,
      "tasks entity": 32314,
      "entity typing": 9656,
      "models popular": 21823,
      "fully leverage": 11956,
      "leverage key": 17750,
      "key properties": 15783,
      "novel algorithm": 23053,
      "algorithm effectively": 1704,
      "model set": 20776,
      "taking advantage": 32037,
      "models method": 21719,
      "model augmented": 20380,
      "augmented data": 2811,
      "additional annotations": 1251,
      "average relative": 3023,
      "improvement 20": 14323,
      "datasets various": 7189,
      "bart t5": 3109,
      "code use": 4827,
      "demonstrate large": 7466,
      "available online": 2991,
      "automatically generate": 2912,
      "exam questions": 10095,
      "work developed": 35694,
      "program synthesis": 26195,
      "learning methods": 17604,
      "methods solve": 20096,
      "problem set": 26011,
      "set questions": 29703,
      "work develop": 35693,
      "compare methods": 5108,
      "problem sets": 26012,
      "set topics": 29711,
      "curate dataset": 6467,
      "dataset benchmark": 6943,
      "benchmark questions": 3406,
      "online code": 23361,
      "code answering": 4721,
      "questions generating": 27112,
      "generating new": 12437,
      "new questions": 22837,
      "perform ablation": 24469,
      "studies comparing": 31263,
      "learning chainofthought": 17550,
      "chainofthought prompting": 4302,
      "prompting using": 26399,
      "using gpt3": 34791,
      "gpt3 opt": 13004,
      "opt codex": 23589,
      "codex chatgpt": 4862,
      "perform best": 24471,
      "highlight transformative": 13637,
      "transformative potential": 33697,
      "models streamline": 22012,
      "significantly reducing": 30084,
      "models chatgpt": 21053,
      "correctness completeness": 6206,
      "critical thinking": 6395,
      "generation problem": 12575,
      "field natural": 11143,
      "trained various": 33434,
      "recipe data": 27630,
      "data present": 6804,
      "application generate": 2129,
      "transformer nonautoregressive": 33737,
      "endtoend speech": 9440,
      "speech recognition": 30787,
      "yield good": 35911,
      "decoder generate": 7247,
      "generate tokens": 12334,
      "computationally inefficient": 5487,
      "speed inference": 30796,
      "designed enable": 7728,
      "parallel generation": 24170,
      "output tokens": 23884,
      "models especially": 21217,
      "challenges improving": 4350,
      "accurately predict": 938,
      "tokens tackle": 33247,
      "tackle challenges": 31995,
      "challenges propose": 4371,
      "propose fast": 26512,
      "number tokens": 23164,
      "glancing language": 12787,
      "model glm": 20552,
      "generates semantic": 12403,
      "semantic embeddings": 29454,
      "ability model": 704,
      "model context": 20442,
      "finally design": 11191,
      "strategy generate": 31122,
      "generate negative": 12306,
      "negative samples": 22663,
      "word error": 35638,
      "error rate": 9715,
      "training improve": 33529,
      "performance experiments": 24593,
      "experiments using": 10496,
      "using public": 34889,
      "task demonstrate": 32106,
      "attain comparable": 2696,
      "performance stateoftheart": 24767,
      "model data": 20448,
      "existing solutions": 10314,
      "heuristic rules": 13532,
      "gpt2 using": 12964,
      "produce new": 26151,
      "new synthetic": 22853,
      "taskspecific knowledge": 32564,
      "knowledge limited": 15879,
      "issue propose": 15661,
      "propose knowledge": 26522,
      "augmentation model": 2805,
      "model pretrained": 20720,
      "novel framework": 23081,
      "framework knowledge": 11873,
      "knowledge single": 15906,
      "utilize knowledge": 34959,
      "task limited": 32153,
      "instances specifically": 15113,
      "tasks unified": 32538,
      "unified texttotext": 34339,
      "objectives different": 23210,
      "different granularity": 8079,
      "knowledge attempt": 15814,
      "multitask training": 22456,
      "experiments synthetic": 10487,
      "data produced": 6809,
      "successfully improves": 31542,
      "performance strong": 24769,
      "strong pretrained": 31189,
      "bert albert": 3495,
      "large margin": 17229,
      "successfully transfers": 31544,
      "task knowledge": 32146,
      "seen unseen": 29367,
      "pretraining work": 25852,
      "past decades": 24388,
      "potential new": 25285,
      "new learning": 22816,
      "learning paradigm": 17626,
      "paradigm nlp": 24158,
      "role data": 28956,
      "process data": 26055,
      "simple principle": 30161,
      "cache large": 3958,
      "large data": 16940,
      "pretraining models": 25822,
      "valuable information": 35011,
      "raw data": 27282,
      "engineering challenges": 9464,
      "models surpass": 22033,
      "surpass strong": 31735,
      "variety nlp": 35068,
      "tasks achieve": 32230,
      "college entrance": 4940,
      "entrance examination": 9659,
      "specifically proposed": 30755,
      "higher average": 13595,
      "addition test": 1248,
      "pretraining natural": 25824,
      "success natural": 31519,
      "unsupervised manner": 34456,
      "manner using": 19551,
      "increasing number": 14626,
      "number models": 23153,
      "data supervised": 6884,
      "showcase superior": 29841,
      "models motivated": 21738,
      "pretraining propose": 25833,
      "propose multitask": 26534,
      "collect largescale": 4918,
      "largescale natural": 17367,
      "11 diverse": 69,
      "soft prompts": 30449,
      "models capacity": 21044,
      "perform specific": 24504,
      "specific task": 30718,
      "model seen": 20770,
      "instruction tuning": 15186,
      "demonstrated effectiveness": 7518,
      "model number": 20659,
      "performance 13": 24513,
      "dataset chinese": 6948,
      "chinese language": 4545,
      "task demands": 32105,
      "general knowledge": 12170,
      "language paper": 16768,
      "dataset named": 7015,
      "model generation": 20549,
      "model produces": 20730,
      "descriptions generated": 7687,
      "order assess": 23669,
      "assess performance": 2601,
      "models task": 22046,
      "chatgpt chatglm": 4460,
      "test results": 32781,
      "reveal current": 28794,
      "language acquisition": 16034,
      "similar natural": 30111,
      "allows obtain": 1815,
      "representation linguistic": 28142,
      "using external": 34776,
      "statistical analysis": 31018,
      "analysis pretrained": 1949,
      "models widely": 22129,
      "nlu natural": 22971,
      "tasks making": 32417,
      "used downstream": 34596,
      "downstream applications": 8673,
      "linguistic theory": 18023,
      "english models": 9486,
      "information language": 14875,
      "models process": 21855,
      "linguistic information": 18013,
      "stages training": 30861,
      "training language": 33537,
      "demonstrate capabilities": 7438,
      "various levels": 35110,
      "fail tasks": 10908,
      "introduce opensource": 15531,
      "opensource framework": 23503,
      "compatible transformerbased": 5206,
      "sensitivity analysis": 29521,
      "architectures bert": 2460,
      "financial sentiment": 11220,
      "word embedding": 35636,
      "learning techniques": 17662,
      "potential applications": 25239,
      "financial sector": 11219,
      "like gpt": 17865,
      "works methods": 35817,
      "investigate performance": 15590,
      "finetuning performance": 11482,
      "performance based": 24528,
      "batch size": 3287,
      "size learning": 30259,
      "learning rate": 17640,
      "gpt2 stable": 12954,
      "layers gpt2": 17439,
      "pattern information": 24407,
      "information maintained": 14886,
      "models infer": 21396,
      "representations encode": 28159,
      "rich semantic": 28875,
      "semantic syntactic": 29477,
      "novel neural": 23101,
      "inductive biases": 14739,
      "relational structures": 27868,
      "output representations": 23879,
      "representations pretrained": 28170,
      "models specifically": 22000,
      "specifically model": 30749,
      "model encodes": 20487,
      "distribution demonstrate": 8391,
      "demonstrate model": 7474,
      "artificially generated": 2542,
      "generated datasets": 12351,
      "random token": 27178,
      "token sequences": 33203,
      "leverage pretrained": 17760,
      "pretrained bert": 25630,
      "models encoder": 21199,
      "encoder decoder": 9346,
      "language datasets": 16058,
      "datasets experiments": 7112,
      "encoding different": 9382,
      "different aspects": 8050,
      "models effectively": 21182,
      "finally explore": 11194,
      "explore training": 10601,
      "reasoning models": 27423,
      "models commonsense": 21067,
      "reasoning tasks": 27456,
      "tasks automatic": 32248,
      "valuable task": 35019,
      "methods achieved": 20001,
      "achieved great": 1006,
      "great progress": 13258,
      "information annotated": 14854,
      "limits performance": 17980,
      "performance methods": 24676,
      "annotation work": 2028,
      "work aims": 35667,
      "aims explore": 1665,
      "explore new": 10594,
      "specifically devise": 30736,
      "addition propose": 1245,
      "framework leverages": 11878,
      "detailed analyses": 7833,
      "used datasets": 34595,
      "effectiveness method": 8958,
      "experiments available": 10420,
      "context based": 5884,
      "computational linguistics": 5469,
      "intended meaning": 15368,
      "word sentence": 35649,
      "larger context": 17318,
      "developing efficient": 7943,
      "complex task": 5296,
      "task recent": 32185,
      "models used": 22099,
      "methods including": 20049,
      "including machine": 14503,
      "learning algorithms": 17534,
      "google t5": 12830,
      "presented training": 25569,
      "training run": 33603,
      "different context": 8058,
      "context lengths": 5899,
      "transformers ability": 33775,
      "ability pretrained": 712,
      "factual knowledge": 10886,
      "knowledge essential": 15847,
      "feedforward networks": 11077,
      "introduce extra": 15508,
      "memory slots": 19829,
      "extra knowledge": 10741,
      "original pretrained": 23717,
      "model train": 20833,
      "modeling ability": 20886,
      "ability original": 706,
      "verify strong": 35219,
      "strong ability": 31160,
      "closedbook question": 4683,
      "answering datasets": 2060,
      "representative tasks": 28187,
      "summarization machine": 31618,
      "translation thoroughly": 33857,
      "thoroughly analyze": 33075,
      "keys values": 15792,
      "way finally": 35431,
      "knowledge stored": 15911,
      "cognitive processes": 4880,
      "powered large": 25332,
      "research understand": 28367,
      "decisionmaking processes": 7238,
      "qualitative study": 26935,
      "study shed": 31397,
      "shed light": 29794,
      "bias language": 3648,
      "model align": 20361,
      "varying degrees": 35171,
      "complex ways": 5298,
      "various criteria": 35083,
      "writing process": 35855,
      "higher levels": 13599,
      "qualitative analysis": 26931,
      "analysis using": 1976,
      "process model": 26074,
      "model writing": 20875,
      "propose theoretical": 26574,
      "causal language": 4240,
      "models general": 21303,
      "task followed": 32126,
      "directions future": 8228,
      "spoken dialogue": 30814,
      "dialogue agents": 8010,
      "realtime feedback": 27331,
      "features pretrained": 11036,
      "pretrained speech": 25750,
      "representation model": 28146,
      "propose metrics": 26531,
      "train evaluate": 33361,
      "metrics vastly": 20151,
      "common approach": 5003,
      "algorithm faster": 1706,
      "deep models": 7335,
      "learning different": 17559,
      "different kinds": 8086,
      "deep networks": 7336,
      "need different": 22624,
      "process inefficient": 26064,
      "consistently improve": 5749,
      "improve model": 14275,
      "training speed": 33620,
      "propose adaptive": 26492,
      "develop new": 7918,
      "problems deep": 26023,
      "lower bound": 19284,
      "surpasses corresponding": 31741,
      "vision language": 35302,
      "rl tasks": 28903,
      "gpt2 transformerxl": 12962,
      "half training": 13366,
      "training cost": 33460,
      "achieve higher": 967,
      "higher comparable": 13596,
      "shows great": 29926,
      "large range": 17274,
      "32k code": 327,
      "code released": 4806,
      "used multiple": 34615,
      "multiple popular": 22409,
      "popular deep": 25115,
      "phishing detection": 24932,
      "phishing emails": 24933,
      "empirical analysis": 9219,
      "leverage knowledge": 17751,
      "knowledge training": 15914,
      "detection task": 7883,
      "task use": 32206,
      "network model": 22698,
      "performance significantly": 24752,
      "indicating effectiveness": 14702,
      "mitigate effect": 20250,
      "use gpt2": 34533,
      "outperforms current": 23816,
      "models f1": 21252,
      "additionally analysis": 1271,
      "analysis individual": 1929,
      "detecting phishing": 7860,
      "question generation": 27069,
      "study investigates": 31348,
      "generated questions": 12382,
      "triples knowledge": 33915,
      "plms typically": 25052,
      "typically trained": 34082,
      "trained natural": 33414,
      "proven effective": 26674,
      "effective lowresource": 8882,
      "effectively utilize": 8935,
      "address challenges": 1313,
      "generate questions": 12315,
      "handle complex": 13407,
      "secondly propose": 29334,
      "trained largescale": 33406,
      "largescale unsupervised": 17388,
      "performance especially": 24583,
      "especially lowresource": 9743,
      "lowresource settings": 19317,
      "settings furthermore": 29736,
      "inference finetuning": 14778,
      "large models": 17233,
      "models nlp": 21754,
      "benefit using": 3484,
      "llms 100": 18398,
      "100 billion": 34,
      "using models": 34844,
      "cases llms": 4204,
      "llms used": 19022,
      "requires access": 28245,
      "weights attention": 35502,
      "attention logits": 2723,
      "consumer gpus": 5818,
      "interactive llm": 15394,
      "llm applications": 18268,
      "models allowing": 20975,
      "allowing train": 1807,
      "based efficient": 3153,
      "finetuning methods": 11453,
      "news corpus": 22878,
      "pretrained autoregressive": 25627,
      "model paper": 20681,
      "shared task": 29784,
      "automatic detection": 2879,
      "using t5": 34925,
      "t5 pretrained": 31960,
      "model iteratively": 20595,
      "consider different": 5702,
      "causal relationships": 4247,
      "model conditioned": 20436,
      "despite training": 7822,
      "small dataset": 30338,
      "approach achieved": 2227,
      "achieved competitive": 1004,
      "competitive performance": 5223,
      "achieves similar": 1065,
      "similar results": 30116,
      "results automatic": 28572,
      "methods automatic": 20008,
      "potential impact": 25261,
      "learning recently": 17643,
      "recently large": 27606,
      "german language": 12731,
      "learning based": 17544,
      "promise improve": 26275,
      "models reliably": 21922,
      "sentences combined": 29552,
      "performed better": 24828,
      "2022 shared": 230,
      "task text": 32198,
      "text complexity": 32830,
      "data best": 6624,
      "gradientbased tuning": 13195,
      "sequencetosequence models": 29620,
      "learning performance": 17628,
      "linguistic tasks": 18022,
      "tasks huge": 32355,
      "cost training": 6255,
      "training larger": 33549,
      "prohibitively expensive": 26240,
      "expensive motivating": 10361,
      "study efficient": 31324,
      "hyperparameter optimization": 13959,
      "hyperparameters training": 13961,
      "apply simple": 2212,
      "simple general": 30150,
      "efficiency performance": 9010,
      "performance gains": 24606,
      "translation natural": 33837,
      "method generalizes": 19923,
      "language pairs": 16767,
      "pretraining improve": 25800,
      "tasks learning": 32398,
      "learning multiple": 17618,
      "global learning": 12796,
      "training improves": 33530,
      "learning unified": 17666,
      "transformers shown": 33795,
      "shown remarkable": 29911,
      "task multitask": 32163,
      "especially natural": 9745,
      "attempts train": 2709,
      "transformers different": 33779,
      "different domains": 8072,
      "usually clear": 34945,
      "study multitask": 31368,
      "learning works": 17672,
      "tasks significantly": 32504,
      "significantly different": 30046,
      "tasks domains": 32302,
      "python code": 26903,
      "carried extensive": 4187,
      "using popular": 34875,
      "training strategies": 33624,
      "finetuning evaluate": 11398,
      "metrics measure": 20142,
      "measure performance": 19735,
      "performance various": 24797,
      "considerable challenges": 5706,
      "challenges models": 4362,
      "finetuning strategy": 11539,
      "learning performs": 17629,
      "performs tasks": 24857,
      "tasks keeping": 32384,
      "accelerating transformerbased": 802,
      "transformerbased text": 33768,
      "generation transformer": 12625,
      "learning language": 17590,
      "models generative": 21315,
      "transformer gpt": 33718,
      "remarkable performance": 28044,
      "generation natural": 12559,
      "processing large": 26107,
      "large input": 16956,
      "context summarization": 5921,
      "single word": 30230,
      "word time": 35651,
      "parallel processing": 24171,
      "significantly degrades": 30044,
      "degrades generation": 7383,
      "high latency": 13571,
      "model inference": 20580,
      "low latency": 19269,
      "high throughput": 13585,
      "summarization generation": 31612,
      "model parallelism": 20686,
      "instructions provide": 15266,
      "operations endtoend": 23569,
      "xilinx alveo": 35871,
      "alveo u280": 1861,
      "high bandwidth": 13552,
      "bandwidth memory": 3090,
      "memory hbm": 19818,
      "hardware efficiency": 13429,
      "energy efficiency": 9448,
      "promising solution": 26297,
      "cloud datacenters": 4703,
      "llms training": 19007,
      "recent large": 27523,
      "llms demonstrated": 18512,
      "demonstrated remarkable": 7541,
      "growing array": 13311,
      "array tasks": 2510,
      "highstakes domains": 13712,
      "address need": 1347,
      "framework leveraging": 11880,
      "knowledge learned": 15877,
      "llms build": 18454,
      "use llms": 34547,
      "compared llms": 5147,
      "llms explore": 18592,
      "embeddings llm": 9143,
      "decision tree": 7234,
      "outperform larger": 23776,
      "6billion parameter": 491,
      "gptj model": 13150,
      "model despite": 20464,
      "despite having": 7783,
      "study generate": 31335,
      "generate interesting": 12295,
      "data code": 6634,
      "code using": 4829,
      "available github": 2978,
      "lms trained": 19117,
      "raw texts": 27285,
      "physical world": 24941,
      "trained text": 33432,
      "cooccurrence statistics": 6128,
      "lms smaller": 19109,
      "models scaled": 21952,
      "larger language": 17321,
      "llms palm": 18836,
      "query llms": 27029,
      "llms typical": 19013,
      "llms significantly": 18949,
      "outperform smaller": 23785,
      "human judgments": 13834,
      "certain types": 4282,
      "short texts": 29819,
      "paper explores": 24049,
      "transformer language": 33723,
      "intrinsic extrinsic": 15492,
      "short text": 29818,
      "scientific publications": 29254,
      "compare results": 5114,
      "results obtained": 28652,
      "different methods": 8102,
      "model yields": 20876,
      "particularly promising": 24353,
      "performance model": 24680,
      "news stories": 22887,
      "represent text": 28135,
      "scientific abstracts": 29248,
      "challenges evaluating": 4341,
      "simple strategy": 30164,
      "transfer new": 33682,
      "given natural": 12756,
      "language prompt": 16806,
      "perform task": 24505,
      "additional training": 1266,
      "prompt cause": 26310,
      "variations model": 35049,
      "model predictions": 20716,
      "task mitigate": 32158,
      "high degree": 13563,
      "effort involved": 9078,
      "lead high": 17464,
      "high quality": 13579,
      "prompting strategy": 26396,
      "proposed prompting": 26617,
      "prompting method": 26388,
      "effective prompt": 8891,
      "questionanswering qa": 27085,
      "prompts encourage": 26414,
      "openended generation": 23457,
      "tend outperform": 32707,
      "model outputs": 20677,
      "true false": 33924,
      "task inputs": 32140,
      "true label": 33925,
      "prompts different": 26411,
      "produce final": 26144,
      "final predictions": 11181,
      "opensource model": 23527,
      "bloom opt": 3790,
      "t0 model": 31933,
      "average performance": 3022,
      "fewshot baseline": 11096,
      "strategy enables": 31118,
      "model match": 20635,
      "match exceed": 19641,
      "exceed performance": 10147,
      "popular benchmarks": 25114,
      "averaged tasks": 3028,
      "outperforms fewshot": 23823,
      "llms shown": 18938,
      "shown exceptional": 29875,
      "exceptional performance": 10170,
      "performance variety": 24793,
      "variety natural": 35065,
      "language tasks": 16830,
      "tasks capabilities": 32257,
      "finetuned llms": 11339,
      "llms indepth": 18715,
      "description generation": 7683,
      "training procedures": 33591,
      "understanding llms": 34246,
      "llms pretrained": 18864,
      "language corpora": 16054,
      "tasks instance": 32376,
      "compared models": 5152,
      "trained exclusively": 33398,
      "dataset finetuned": 6992,
      "benchmark llms": 3396,
      "llms successfully": 18978,
      "successfully complete": 31537,
      "data compared": 6646,
      "compared previous": 5162,
      "model llms": 20628,
      "llms evaluate": 18570,
      "t5based models": 31974,
      "encoderdecoder architecture": 9363,
      "promote research": 26301,
      "research llms": 28335,
      "opensource largescale": 23514,
      "dataset distilled": 6975,
      "strong language": 31177,
      "models incur": 21394,
      "selfattention mechanism": 29409,
      "work proposes": 35762,
      "time memory": 33134,
      "memory complexity": 19805,
      "simple alternative": 30140,
      "outperforms prior": 23847,
      "prior methods": 25937,
      "retains 99": 28721,
      "generation challenging": 12470,
      "endtoend neural": 9437,
      "neural methods": 22735,
      "methods require": 20087,
      "require substantial": 28226,
      "substantial training": 31478,
      "data realworld": 6825,
      "issues access": 15665,
      "examples different": 10121,
      "different domain": 8071,
      "domain schema": 8592,
      "new approach": 22776,
      "diverse settings": 8462,
      "efficient use": 9065,
      "consists steps": 5766,
      "steps data": 31058,
      "sentence fusion": 29535,
      "offtheshelf pretrained": 23331,
      "finetuning data": 11384,
      "model understand": 20848,
      "stage uses": 30859,
      "like t5": 17895,
      "datasets different": 7097,
      "different scenarios": 8136,
      "scenarios including": 29209,
      "generalization unseen": 12228,
      "outofdomain data": 23751,
      "data experimental": 6702,
      "consistently achieves": 5744,
      "achieves significant": 1062,
      "improvement baselines": 14332,
      "dataset zeroshot": 7057,
      "zeroshot setting": 35994,
      "models understanding": 22094,
      "nlp research": 22945,
      "models abilities": 20931,
      "complex reasoning": 5289,
      "reasoning abilities": 27368,
      "abilities work": 648,
      "focuses simple": 11668,
      "end introduce": 9412,
      "questionanswering dataset": 27083,
      "dataset involving": 7006,
      "binary classification": 3732,
      "questions mcq": 27123,
      "models gpt3": 21327,
      "gpt3 gpt2": 12995,
      "gpt2 t5": 12956,
      "struggle answer": 31238,
      "questions correctly": 27103,
      "gpt3 achieves": 12981,
      "relevant knowledge": 27943,
      "answer question": 2050,
      "additional knowledge": 1257,
      "performance overall": 24708,
      "performance remains": 24736,
      "models reason": 21898,
      "models finetuning": 21279,
      "models collection": 21061,
      "collection datasets": 4928,
      "instructions shown": 15273,
      "shown improve": 29889,
      "performance generalization": 24611,
      "instruction finetuning": 15157,
      "finetuning particular": 11474,
      "focus scaling": 11656,
      "scaling number": 29178,
      "tasks scaling": 32492,
      "size finetuning": 30249,
      "data instruction": 6739,
      "model classes": 20419,
      "evaluation benchmarks": 9926,
      "benchmarks mmlu": 3463,
      "mmlu bbh": 20313,
      "generation instance": 12522,
      "flanpalm 540b": 11590,
      "tasks outperforms": 32439,
      "outperforms palm": 23842,
      "palm 540b": 23990,
      "performance benchmarks": 24531,
      "publicly release": 26862,
      "achieve strong": 993,
      "strong fewshot": 31170,
      "fewshot performance": 11117,
      "compared larger": 5143,
      "models palm": 21789,
      "finetuning general": 11408,
      "general method": 12179,
      "evaluation large": 9965,
      "knowledge encoded": 15841,
      "encoded pretrained": 9343,
      "introduce benchmark": 15499,
      "sentence pairs": 29539,
      "mandarin chinese": 19534,
      "syntactic semantic": 31823,
      "minimal pairs": 20191,
      "english blimp": 9475,
      "data generation": 6722,
      "generation process": 12576,
      "process test": 26086,
      "available pretrained": 2996,
      "pretrained monolingual": 25731,
      "average accuracy": 3009,
      "achieves highest": 1048,
      "highest accuracy": 13619,
      "lms larger": 19094,
      "larger ones": 17334,
      "lms strong": 19112,
      "gender number": 12151,
      "perform better": 24472,
      "use multiple": 34553,
      "multiple nodes": 22406,
      "step contrast": 31039,
      "improves accuracy": 14369,
      "accuracy distribution": 869,
      "distribution shift": 8396,
      "compared baseline": 5122,
      "opt language": 23594,
      "models 13b": 20922,
      "13b parameters": 132,
      "common crawl": 5006,
      "enables finetuning": 9300,
      "finetuning settings": 11519,
      "gpu hours": 13170,
      "scale increasing": 29134,
      "100b parameters": 49,
      "models increasingly": 21390,
      "emergent capabilities": 9184,
      "capabilities given": 4023,
      "openaccess multilingual": 23429,
      "multilingual language": 22310,
      "identify architecture": 14004,
      "training setup": 33614,
      "best use": 3580,
      "ablation study": 737,
      "comparing different": 5189,
      "zeroshot generalization": 35977,
      "impact various": 14142,
      "various popular": 35133,
      "study performance": 31372,
      "performance multilingual": 24682,
      "scaling behaviour": 29160,
      "target model": 32055,
      "models code": 21056,
      "code opensourced": 4795,
      "semiparametric language": 29499,
      "require huge": 28219,
      "number model": 23151,
      "solving multiple": 30514,
      "multiple natural": 22403,
      "settings addition": 29729,
      "costly model": 6264,
      "model retraining": 20759,
      "paper develop": 24033,
      "develop novel": 7920,
      "novel semiparametric": 23107,
      "causality knowledge": 4250,
      "input instance": 15013,
      "knowledge type": 15916,
      "knowledge augmentation": 15815,
      "generate output": 12309,
      "answer input": 2045,
      "input output": 15018,
      "model knowledge": 20598,
      "plays role": 25030,
      "superior zeroshot": 31658,
      "zeroshot performance": 35987,
      "performance unseen": 24790,
      "outperforms large": 23827,
      "emergent abilities": 9182,
      "abilities smaller": 639,
      "smaller model": 30381,
      "scale compared": 29129,
      "models leveraging": 21445,
      "leveraging pretrained": 17792,
      "failure analysis": 10917,
      "recently gained": 27598,
      "domain text": 8599,
      "text summarization": 32954,
      "generation questionanswering": 12586,
      "models long": 21687,
      "long short": 19179,
      "short term": 29816,
      "term memory": 32730,
      "memory lstm": 19821,
      "leverage attention": 17744,
      "pretrained causal": 25633,
      "model downstream": 20473,
      "downstream task": 8685,
      "task generating": 32132,
      "semiconductor industry": 29495,
      "compare different": 5105,
      "generative task": 12705,
      "task observe": 32169,
      "gpt2 outperformed": 12933,
      "gpt2 trained": 12959,
      "bert bart": 3497,
      "evaluation metric": 9974,
      "better evaluation": 3601,
      "human judgment": 13833,
      "existing metrics": 10296,
      "carbon footprint": 4160,
      "bloom 176b": 3782,
      "parameter language": 24184,
      "comes cost": 4970,
      "given training": 12779,
      "ml models": 20290,
      "significant computational": 29967,
      "aim quantify": 1648,
      "final training": 11186,
      "carbon emissions": 4159,
      "deployment inference": 7648,
      "user queries": 34667,
      "conclude discussion": 5551,
      "discussion regarding": 8301,
      "footprint ml": 11713,
      "models future": 21299,
      "research directions": 28309,
      "contribute improving": 6029,
      "ability reason": 719,
      "raw text": 27284,
      "text ability": 32813,
      "combine multiple": 4957,
      "multiple evidence": 22391,
      "evidence propose": 10062,
      "novel learning": 23091,
      "helps language": 13524,
      "better understand": 3633,
      "perform complex": 24476,
      "compositional reasoning": 5328,
      "reasoning model": 27422,
      "model learns": 20609,
      "multihop question": 22287,
      "comprehension model": 5345,
      "model predict": 20714,
      "predict answer": 25405,
      "using language": 34804,
      "outperform baseline": 23764,
      "absolute f1": 762,
      "task report": 32187,
      "make sentences": 19483,
      "annotated human": 2022,
      "respectively demonstrate": 28457,
      "difficult task": 8173,
      "task zeroshot": 32210,
      "given limitations": 12752,
      "limitations approaches": 17915,
      "generation method": 12545,
      "transformers scratch": 33794,
      "finetune t5": 11302,
      "improved finetuning": 14310,
      "dataset derived": 6972,
      "model large": 20602,
      "shown able": 29867,
      "able perform": 749,
      "perform new": 24497,
      "tasks based": 32250,
      "language instructions": 16098,
      "instructions capabilities": 15245,
      "led widespread": 17688,
      "adoption llms": 1412,
      "llms developed": 18538,
      "present bloom": 25516,
      "model designed": 20462,
      "decoderonly transformer": 7265,
      "corpus dataset": 6179,
      "dataset comprising": 6956,
      "programming languages": 26200,
      "performance wide": 24813,
      "wide variety": 35566,
      "stronger results": 31202,
      "multitask prompted": 22453,
      "prompted finetuning": 26367,
      "research applications": 28291,
      "using llms": 34825,
      "llms publicly": 18882,
      "release models": 27912,
      "responsible ai": 28520,
      "working memory": 35804,
      "llms led": 18752,
      "generation abilities": 12449,
      "massive amounts": 19623,
      "pretraining downstream": 25794,
      "applications provide": 2173,
      "factual information": 10885,
      "information presented": 14899,
      "context remains": 5912,
      "remains explored": 27995,
      "behavior llm": 3318,
      "context contains": 5885,
      "knowledge enables": 15840,
      "enables model": 9304,
      "specific model": 30706,
      "internal knowledge": 15437,
      "context llms": 5901,
      "llms demonstrate": 18508,
      "demonstrate stateoftheart": 7497,
      "stateoftheart t5": 30996,
      "pretrained finetuned": 25643,
      "exhibit poor": 10225,
      "poor controllability": 25102,
      "size solution": 30285,
      "solution propose": 30477,
      "standard supervised": 30882,
      "supervised datasets": 31672,
      "comprehensive evaluation": 5367,
      "architectures sizes": 2472,
      "generalization gap": 12215,
      "learning pretrained": 17631,
      "pretrained large": 25697,
      "shown great": 29879,
      "tasks exhibit": 32319,
      "exhibit low": 10224,
      "various nlp": 35130,
      "tasks just": 32383,
      "solve task": 30497,
      "finetuning known": 11423,
      "known incontext": 15935,
      "learning work": 17671,
      "work look": 35734,
      "indistribution id": 14714,
      "outofdistribution ood": 23749,
      "models semantic": 21961,
      "parsing tasks": 24321,
      "tasks incontext": 32369,
      "train model": 33367,
      "families opt": 10970,
      "opt bloom": 23588,
      "codegen codex": 4839,
      "different number": 8114,
      "gap models": 12097,
      "information overload": 14895,
      "major obstacle": 19443,
      "explosive growth": 10626,
      "growth scientific": 13322,
      "scientific literature": 29252,
      "useful insights": 34640,
      "scientific knowledge": 29251,
      "search engines": 29309,
      "train large": 33364,
      "material knowledge": 19663,
      "technical knowledge": 32606,
      "despite trained": 7821,
      "trained general": 33400,
      "outperforms bloom": 23811,
      "demonstrate potential": 7480,
      "model benefit": 20394,
      "scientific community": 29249,
      "use search": 34567,
      "algorithms possible": 1722,
      "identify mentions": 14011,
      "uses texttotext": 34716,
      "seq2seq paradigm": 29593,
      "underlying language": 34155,
      "obtain stateoftheart": 23253,
      "stateoftheart accuracy": 30920,
      "data sets": 6860,
      "sets experiments": 29718,
      "experiments zeroshot": 10505,
      "fewshot setting": 11122,
      "setting using": 29727,
      "available training": 3002,
      "languages previous": 16906,
      "previous approaches": 25865,
      "approaches significantly": 2393,
      "exceed previous": 10148,
      "previous supervised": 25887,
      "supervised stateoftheart": 31690,
      "results tested": 28696,
      "adapting large": 1208,
      "backpropagation finetuning": 3068,
      "models different": 21154,
      "methods reduce": 20084,
      "reduce number": 27721,
      "parameters require": 24285,
      "require gradientbased": 28218,
      "novel approach": 23056,
      "model adaptation": 20355,
      "model demonstrate": 20455,
      "demonstrate simple": 7496,
      "lora parameters": 19233,
      "model fewshot": 20519,
      "fewshot examples": 11102,
      "multitask finetuning": 22446,
      "finetuning mtf": 11456,
      "diverse language": 8435,
      "tasks evaluate": 32316,
      "datasets effectively": 7100,
      "flexible efficient": 11617,
      "efficient way": 9066,
      "models diverse": 21164,
      "diverse downstream": 8426,
      "applications understanding": 2179,
      "study diverse": 31320,
      "landscape large": 16023,
      "llms lens": 18754,
      "bloom model": 3788,
      "performance bloom": 24535,
      "decoderonly llms": 7261,
      "llms compared": 18475,
      "models achieve": 20944,
      "model variants": 20858,
      "performance does": 24571,
      "does scale": 8541,
      "parameter size": 24198,
      "unlike llms": 34396,
      "llms like": 18756,
      "experiments finetuning": 10443,
      "bloom models": 3789,
      "performs similarly": 24856,
      "similarly better": 30134,
      "zeroshot crosslingual": 35964,
      "gpt3 models": 13003,
      "models scientific": 21954,
      "paper examines": 24043,
      "typically focus": 34078,
      "limited set": 17965,
      "high similarity": 13583,
      "realistic setup": 27317,
      "dataset containing": 6962,
      "using dataset": 34763,
      "lms stateoftheart": 19111,
      "stateoftheart lms": 30950,
      "tasks highlighting": 32354,
      "challenges posed": 4368,
      "demonstrated substantial": 7550,
      "substantial gains": 31465,
      "largelanguage models": 17304,
      "paper evaluate": 24040,
      "reasoning benchmark": 27379,
      "tasks aim": 32239,
      "examine performance": 10104,
      "performance smaller": 24756,
      "smaller models": 30383,
      "llama2 mpt": 18189,
      "mpt falcon": 22246,
      "competitive accuracy": 5218,
      "results using": 28702,
      "understand model": 34196,
      "performance finally": 24600,
      "finally conduct": 11189,
      "robustness tests": 28950,
      "tests using": 32809,
      "using various": 34937,
      "various methods": 35116,
      "performance numerous": 24698,
      "knowledge large": 15871,
      "word cooccurrence": 35635,
      "corpora contain": 6164,
      "llms trained": 19006,
      "trained predict": 33418,
      "words context": 35656,
      "performance diverse": 24569,
      "semantic tasks": 29478,
      "tasks requiring": 32485,
      "important understudied": 14216,
      "question llms": 27072,
      "llms semantic": 18931,
      "knowledge common": 15825,
      "assign higher": 2623,
      "llms possess": 18857,
      "models particular": 21805,
      "teacher llms": 32583,
      "llms consistent": 18485,
      "active vs": 1148,
      "vs passive": 35402,
      "llm representations": 18358,
      "results important": 28628,
      "important aspects": 14197,
      "linguistic patterns": 18016,
      "highlight gap": 13629,
      "memory transformer": 19833,
      "processing long": 26112,
      "long documents": 19172,
      "transformer variants": 33744,
      "stateoftheart different": 30930,
      "summarization paper": 31621,
      "fine tuning": 11265,
      "study aims": 31299,
      "ability proposed": 717,
      "model handle": 20564,
      "used t5": 34628,
      "t5 transformer": 31966,
      "modeling task": 20907,
      "specific training": 30721,
      "parameters ablation": 24220,
      "study reveals": 31390,
      "ability using": 728,
      "degradation performance": 7376,
      "impact language": 14126,
      "characteristics multilingual": 4427,
      "mbert xlmr": 19712,
      "xlmr mt5": 35876,
      "mt5 mbart": 22255,
      "enabling natural": 9326,
      "transfer highresource": 33673,
      "understand models": 34197,
      "crosslingual signals": 6418,
      "texts language": 32996,
      "model appears": 20369,
      "implicitly learn": 14180,
      "raises questions": 27170,
      "source target": 30569,
      "model statistical": 20808,
      "crosslingual understanding": 6422,
      "source language": 30562,
      "key finding": 15768,
      "finding work": 11228,
      "syntax morphology": 31832,
      "lexical similarity": 17803,
      "given language": 12751,
      "able predict": 750,
      "scale number": 29146,
      "language data": 16056,
      "data points": 6797,
      "knowledge generative": 15854,
      "important role": 14210,
      "sequential decisionmaking": 29626,
      "decisionmaking problems": 7236,
      "knowledge required": 15902,
      "largescale generative": 17353,
      "models glms": 21320,
      "textual outputs": 33035,
      "decisionmaking propose": 7239,
      "algorithm named": 1711,
      "task goal": 32135,
      "knowledge proposed": 15893,
      "fills gap": 11169,
      "accordingly propose": 849,
      "glm based": 12791,
      "everyday tasks": 10058,
      "secure multiparty": 29346,
      "multiparty computation": 22375,
      "approach generating": 2289,
      "given news": 12759,
      "news story": 22888,
      "summarization task": 31625,
      "news article": 22875,
      "task produce": 32180,
      "corpora model": 6168,
      "results showcase": 28677,
      "data efficiency": 6688,
      "quality training": 26982,
      "training efficiency": 33507,
      "efficiency efficient": 9003,
      "efficient data": 9030,
      "recent advances": 27499,
      "root causes": 28971,
      "speed model": 30797,
      "rapidly evolving": 27255,
      "efficiently use": 9076,
      "use training": 34576,
      "foundation model": 11796,
      "framework focuses": 11856,
      "end present": 9415,
      "work achieves": 35663,
      "95 model": 603,
      "quality compared": 26947,
      "better model": 3611,
      "benefit additional": 3480,
      "including gpt3": 14484,
      "case study": 4196,
      "able capture": 742,
      "native speakers": 22500,
      "earlier results": 8771,
      "results showing": 28681,
      "examine effects": 10100,
      "effects gender": 8979,
      "study examines": 31329,
      "able produce": 751,
      "human raters": 13858,
      "generated model": 12374,
      "model explain": 20502,
      "sentence simplification": 29544,
      "simplification using": 30175,
      "using transformers": 34934,
      "maintaining original": 19429,
      "research focused": 28317,
      "focused tackling": 11664,
      "external linguistic": 10732,
      "sentences paper": 29557,
      "models experiment": 21234,
      "combination gpt2": 4950,
      "achieving best": 1085,
      "sari score": 29102,
      "mechanical turk": 19746,
      "dataset significantly": 7037,
      "better previous": 3619,
      "results code": 28578,
      "multilingual large": 22313,
      "multilingual dataset": 22305,
      "dataset used": 7051,
      "models datasets": 21124,
      "datasets analysis": 7064,
      "range research": 27210,
      "distributed training": 8388,
      "training paper": 33582,
      "collaborative research": 4907,
      "takes step": 32034,
      "diversity tasks": 8484,
      "main goal": 19397,
      "share lessons": 29781,
      "lessons learned": 17723,
      "impact social": 14137,
      "large deep": 16943,
      "different contexts": 8060,
      "sparsely activated": 30623,
      "computation costs": 5445,
      "models efficient": 21184,
      "terms quality": 32750,
      "computation cost": 5444,
      "models remain": 21925,
      "scratch large": 29291,
      "large scale": 17275,
      "way reuse": 35447,
      "training costs": 33461,
      "model dense": 20459,
      "base large": 3119,
      "large xl": 17302,
      "models vision": 22117,
      "models respectively": 21936,
      "respectively significantly": 28464,
      "dense counterparts": 7605,
      "using 50": 34724,
      "computation budget": 5443,
      "syntactic evaluations": 31815,
      "ask models": 2553,
      "models stable": 22007,
      "make judgements": 19471,
      "just single": 15734,
      "input does": 15008,
      "match language": 19643,
      "models training": 22077,
      "input sentences": 15026,
      "robust models": 28936,
      "contexts paper": 5945,
      "properties input": 26473,
      "context length": 5898,
      "syntactic phenomena": 31821,
      "randomly sampled": 27182,
      "linguistic contexts": 18009,
      "tested models": 32796,
      "gpt2 variants": 12965,
      "variants opt": 35043,
      "unrelated inputs": 34426,
      "changes model": 4411,
      "matching context": 19659,
      "test inputs": 32772,
      "lexical overlap": 17800,
      "highly specific": 13669,
      "explained models": 10532,
      "learning abilities": 17530,
      "shown perform": 29902,
      "investigate hypothesis": 15584,
      "ability large": 689,
      "model incontext": 20578,
      "attention heads": 2719,
      "feed forward": 11055,
      "task performance": 32173,
      "performance substantial": 24771,
      "number incontext": 23143,
      "incontext examples": 14544,
      "small set": 30367,
      "score highly": 29269,
      "ability perform": 710,
      "induction heads": 14736,
      "overall study": 23915,
      "provides insights": 26757,
      "perform incontext": 24491,
      "learning multilingual": 17614,
      "multilingual sequencetosequence": 22328,
      "nlp recent": 22943,
      "nlp large": 22934,
      "increased model": 14609,
      "size large": 30255,
      "data despite": 6675,
      "languages additionally": 16863,
      "focused encoderonly": 11661,
      "tasks named": 32423,
      "paper argue": 24013,
      "generative architectures": 12652,
      "leverage powerful": 17759,
      "multilingual pretrained": 22323,
      "pretrained sequencetosequence": 25748,
      "models mt5": 21739,
      "eliminating need": 9110,
      "need specialized": 22642,
      "using approach": 34732,
      "approach experiments": 2275,
      "previously published": 25901,
      "results existing": 28607,
      "benchmarks results": 3471,
      "tuning language": 33985,
      "tuning enables": 33974,
      "enables pretrained": 9306,
      "language descriptions": 16061,
      "approaches rely": 2388,
      "rely vast": 27973,
      "vast amounts": 35181,
      "human supervision": 13868,
      "user interactions": 34658,
      "instructions large": 15256,
      "large dataset": 16941,
      "diverse instructions": 8434,
      "examples instructions": 10130,
      "noise training": 22985,
      "training opensource": 33580,
      "surpassing performance": 31758,
      "models t0": 22040,
      "various benchmarks": 35077,
      "modelgenerated data": 20883,
      "personality traits": 24885,
      "explored aspects": 10608,
      "human personality": 13852,
      "used assess": 34584,
      "work explore": 35704,
      "explore question": 10600,
      "models exhibited": 21229,
      "generation example": 12495,
      "highly predictable": 13665,
      "dialog systems": 8005,
      "crowdsourced dataset": 6432,
      "human subjects": 13867,
      "chainofthought reasoning": 4304,
      "llms surprisingly": 18986,
      "generating natural": 12435,
      "language reasoning": 16812,
      "reasoning steps": 27454,
      "answering qa": 2069,
      "parameters using": 24294,
      "using question": 34895,
      "retrieve relevant": 28769,
      "relevant text": 27946,
      "llms observe": 18822,
      "using retrieved": 34903,
      "results improve": 28630,
      "substantially improves": 31483,
      "improves retrieval": 14393,
      "gains outofdistribution": 12072,
      "reduces model": 27738,
      "model hallucination": 20562,
      "cot reasoning": 6283,
      "data prompts": 6812,
      "safety large": 29048,
      "tested different": 32795,
      "different llms": 8096,
      "llms using": 19025,
      "instruction finetuned": 15154,
      "gpt35 gpt4": 13021,
      "gpt4 showed": 13117,
      "evaluated llms": 9879,
      "llms gpt": 18643,
      "gpt series": 12863,
      "impact finetuning": 14123,
      "models following": 21292,
      "using direct": 34768,
      "direct preference": 8213,
      "preference optimization": 25469,
      "effectively reduce": 8928,
      "based findings": 3160,
      "metrics evaluate": 20135,
      "improve safety": 14297,
      "safety llms": 29052,
      "labeling tasks": 15962,
      "tasks absence": 32229,
      "readily available": 27302,
      "data given": 6724,
      "task language": 32147,
      "data annotation": 6597,
      "parallel corpora": 24167,
      "leverages large": 17768,
      "models stateoftheart": 22010,
      "stateoftheart machine": 30952,
      "candidates using": 3997,
      "using multilingual": 34847,
      "model generated": 20545,
      "generated candidates": 12344,
      "conducted experiments": 5633,
      "outperforms previous": 23843,
      "wide margin": 35548,
      "lack highquality": 15990,
      "highquality training": 13703,
      "tasks code": 32263,
      "data publicly": 6820,
      "massive language": 19626,
      "models accurately": 20942,
      "pruned oneshot": 26805,
      "gpt family": 12847,
      "family models": 10982,
      "50 sparsity": 423,
      "oneshot retraining": 23351,
      "minimal loss": 20189,
      "loss accuracy": 19241,
      "pruning method": 26813,
      "designed work": 7748,
      "efficiently accurately": 9068,
      "gptfamily models": 13143,
      "models execute": 21224,
      "largest available": 17390,
      "available opensource": 2994,
      "opensource models": 23528,
      "models opt175b": 21778,
      "opt175b bloom176b": 23606,
      "billion weights": 3722,
      "weight quantization": 35496,
      "approaches code": 2370,
      "generative ai": 12645,
      "ai models": 1612,
      "stable diffusion": 30849,
      "perform tasks": 24506,
      "models industry": 21395,
      "ai capable": 1602,
      "texts images": 32993,
      "images like": 14085,
      "model text": 20829,
      "model images": 20568,
      "images text": 14088,
      "texts texts": 33003,
      "like chatgpt": 17853,
      "code like": 4771,
      "like codex": 17858,
      "codex model": 4863,
      "scientific texts": 29260,
      "model create": 20447,
      "algorithms like": 1721,
      "model bert": 20395,
      "albert roberta": 1696,
      "roberta t5": 28922,
      "t5 gpt": 31947,
      "proven promising": 26676,
      "recent nlp": 27538,
      "tuning model": 34000,
      "works proposed": 35824,
      "understanding methods": 34252,
      "methods work": 20111,
      "work paper": 35740,
      "finetuning procedure": 11494,
      "addition able": 1235,
      "theoretical analysis": 33048,
      "analysis framework": 1922,
      "design effective": 7702,
      "novel strategies": 23110,
      "extensively evaluate": 10717,
      "proposed approaches": 26592,
      "used realworld": 34621,
      "classification datasets": 4594,
      "datasets experiment": 7107,
      "experiment results": 10379,
      "results proposed": 28659,
      "generation style": 12607,
      "contextually appropriate": 5969,
      "generation systems": 12610,
      "systems existing": 31897,
      "existing approaches": 10261,
      "textual style": 33039,
      "transfer large": 33675,
      "large volumes": 17301,
      "data argue": 6603,
      "data second": 6852,
      "novel task": 23113,
      "analysis model": 1942,
      "based text": 3230,
      "text generator": 32888,
      "quantitative qualitative": 26994,
      "approach generate": 2287,
      "generic text": 12720,
      "text prompts": 32923,
      "prompts code": 26407,
      "data accessible": 6587,
      "effectiveness large": 8950,
      "models dialog": 21153,
      "increased size": 14612,
      "achieve high": 966,
      "high level": 13572,
      "level performance": 17735,
      "tasks question": 32468,
      "answering summarization": 2075,
      "summarization large": 31614,
      "tasks realm": 32472,
      "llms language": 18741,
      "understanding capabilities": 34213,
      "evaluation task": 10018,
      "prompting llms": 26387,
      "paper shows": 24133,
      "choice datasets": 4551,
      "training model": 33567,
      "task prompt": 32181,
      "datasets model": 7148,
      "evaluation performs": 9987,
      "paper investigates": 24072,
      "examples prompt": 10141,
      "example selection": 10110,
      "affect models": 1538,
      "performance step": 24768,
      "recent innovations": 27521,
      "nvidia a100": 23191,
      "inference stateoftheart": 14811,
      "optimizers like": 23653,
      "attentionbased llms": 2747,
      "llms paper": 18837,
      "light propose": 17833,
      "remains fixed": 27996,
      "automatically identifies": 2919,
      "empirically evaluate": 9249,
      "multiple tasks": 22422,
      "llm finetuning": 18307,
      "sparsity ratios": 30635,
      "generating realistic": 12441,
      "tabular data": 31992,
      "data common": 6645,
      "common form": 5007,
      "multiple models": 22402,
      "available generate": 2977,
      "tabular datasets": 31993,
      "ability produce": 716,
      "data challenging": 6631,
      "challenging requires": 4395,
      "tables introduce": 31988,
      "using autoregressive": 34738,
      "autoregressive gpt2": 2940,
      "model generates": 20546,
      "sequencetosequence seq2seq": 29621,
      "model implement": 20570,
      "using realworld": 34897,
      "baseline model": 3254,
      "outofthebox large": 23758,
      "zeroshot dense": 35966,
      "improve zeroshot": 14302,
      "ability language": 687,
      "memory inference": 19819,
      "model zeroshot": 20877,
      "dense retrieval": 7610,
      "setting augmenting": 29722,
      "strong zeroshot": 31194,
      "retrieval accuracy": 28730,
      "tasks included": 32363,
      "benchmark outperforms": 3404,
      "computation steps": 5450,
      "robust generalization": 28933,
      "parameters plan": 24276,
      "content generation": 5861,
      "complex diverse": 5273,
      "way generating": 35435,
      "generating content": 12416,
      "lack ability": 15974,
      "generate content": 12268,
      "incredibly effective": 14653,
      "trained llms": 33409,
      "llms finetuned": 18610,
      "accelerating training": 801,
      "training new": 33573,
      "tasks introduce": 32378,
      "finetuned gpt2": 11317,
      "generate diverse": 12273,
      "key challenges": 15756,
      "far know": 10988,
      "model combined": 20429,
      "enables generation": 9301,
      "content code": 5853,
      "main challenges": 19391,
      "modern transformer": 22173,
      "architectures like": 2466,
      "like bert": 17845,
      "strengths weaknesses": 31143,
      "language vision": 16854,
      "vision model": 35306,
      "vision models": 35307,
      "present empirical": 25526,
      "empirical evaluation": 9220,
      "evaluation different": 9941,
      "different lms": 8101,
      "gpt2 opt": 12931,
      "benchmark introduce": 3394,
      "unified multitask": 34334,
      "qa datasets": 26908,
      "structured explanations": 31220,
      "question used": 27076,
      "used produce": 34620,
      "extensive evaluation": 10685,
      "evaluation popular": 9990,
      "models fewshot": 21267,
      "models lag": 21414,
      "lag human": 16014,
      "producing structured": 26166,
      "believe work": 3341,
      "work provide": 35763,
      "multistep reasoning": 22442,
      "reasoning conversational": 27398,
      "art large": 2514,
      "large transformerbased": 17281,
      "gpt t5": 12864,
      "deep understanding": 7346,
      "understanding contextual": 34217,
      "semantics language": 29488,
      "enabled significant": 9291,
      "significant advances": 29958,
      "including development": 14471,
      "systems capable": 31888,
      "complete tasks": 5250,
      "levels reasoning": 17741,
      "reasoning humans": 27411,
      "ai research": 1615,
      "approaches include": 2377,
      "ai paper": 1614,
      "used evaluating": 34599,
      "commonsense capabilities": 5035,
      "capabilities stateoftheart": 4073,
      "dialogue models": 8017,
      "aspect based": 2563,
      "based sentiment": 3221,
      "analysis introduce": 1932,
      "aspectbased sentiment": 2570,
      "positive negative": 25194,
      "negative neutral": 22661,
      "examples training": 10144,
      "instruction tune": 15182,
      "tune model": 33958,
      "significant performance": 29999,
      "performance improvements": 24627,
      "sota approaches": 30531,
      "7x larger": 551,
      "models competitive": 21074,
      "strong generalization": 31173,
      "sample efficiency": 29066,
      "train data": 33360,
      "data required": 6837,
      "tuning approaches": 33968,
      "assess quality": 2603,
      "quality instructions": 26970,
      "qa language": 26911,
      "stateoftheart language": 30936,
      "strategies complex": 31102,
      "public data": 26833,
      "data standard": 6880,
      "complex questions": 5287,
      "problems does": 26026,
      "different cultures": 8063,
      "climate change": 4655,
      "knowledge skills": 15907,
      "sensitive data": 29515,
      "data protection": 6815,
      "feedback recent": 11068,
      "chatgpt galactica": 4470,
      "great potential": 13254,
      "limitations llm": 17925,
      "paper start": 24135,
      "research papers": 28341,
      "challenges llm": 4358,
      "llm terms": 18373,
      "evaluation accuracy": 9916,
      "accuracy fairness": 881,
      "discuss challenges": 8291,
      "challenges associated": 4336,
      "including domain": 14473,
      "decomposition efficient": 7290,
      "promising research": 26295,
      "trends using": 33890,
      "patterns training": 24414,
      "prompting strategies": 26395,
      "learning supervised": 17659,
      "guiding large": 13360,
      "prompting introduce": 26379,
      "introduce directional": 15503,
      "llms specific": 18963,
      "llms method": 18799,
      "method employs": 19909,
      "policy model": 25085,
      "guide llms": 13349,
      "llms generating": 18637,
      "direct llm": 8212,
      "model explore": 20504,
      "align llms": 1732,
      "desired behaviors": 7763,
      "using labeled": 34803,
      "data reinforcement": 6829,
      "offline online": 23322,
      "based llms": 3193,
      "llms output": 18835,
      "assess method": 2597,
      "dialogue response": 8018,
      "response generation": 28477,
      "generation chainofthought": 12468,
      "consistently improves": 5751,
      "improves llms": 14382,
      "llms chatgpt": 18463,
      "supervised tasks": 31691,
      "using minimal": 34842,
      "data notably": 6785,
      "using just": 34802,
      "dialogues multiwoz": 8025,
      "multiwoz dataset": 22468,
      "dataset approach": 6939,
      "approach enhances": 2271,
      "chatgpts performance": 4507,
      "performance impressive": 24623,
      "matching surpassing": 19660,
      "fully supervised": 11957,
      "models additionally": 20960,
      "chainofthought prompt": 4301,
      "prompt generated": 26326,
      "generated approach": 12343,
      "approach improves": 2297,
      "reasoning accuracy": 27377,
      "accuracy compared": 865,
      "automatically generated": 2914,
      "models continue": 21107,
      "resources required": 28446,
      "learning leverage": 17598,
      "overhead associated": 23936,
      "associated model": 2651,
      "models computer": 21085,
      "proven challenging": 26672,
      "modern deep": 22157,
      "key value": 15789,
      "successfully implement": 31540,
      "model model": 20644,
      "parameters best": 24229,
      "model date": 20453,
      "generation comprehension": 12474,
      "comprehension natural": 5347,
      "modifying transformer": 22184,
      "transformer block": 33711,
      "self attention": 29404,
      "computational complexity": 5457,
      "linear complexity": 17986,
      "sequence length": 29601,
      "length input": 17706,
      "input tokens": 15035,
      "models tested": 22055,
      "benchmarks maintaining": 3460,
      "fewer operations": 11088,
      "hardware leverage": 13430,
      "past work": 24390,
      "generation settings": 12601,
      "tools work": 33275,
      "democratize access": 7425,
      "extraction models": 10770,
      "decoderonly encoderdecoder": 7254,
      "showcase potential": 29838,
      "gender biases": 12150,
      "good practices": 12822,
      "model evaluations": 20496,
      "llama open": 18133,
      "foundation language": 11792,
      "models introduce": 21404,
      "introduce llama": 15514,
      "llama collection": 18089,
      "models ranging": 21889,
      "7b 65b": 529,
      "65b parameters": 478,
      "trillions tokens": 33910,
      "possible train": 25217,
      "train stateoftheart": 33376,
      "using publicly": 34890,
      "gpt3 175b": 12980,
      "competitive best": 5222,
      "best models": 3565,
      "models research": 21933,
      "currently largest": 6543,
      "largest language": 17394,
      "model explicitly": 20503,
      "search engine": 29308,
      "corpus date": 6180,
      "opensourced available": 23548,
      "available hugging": 2979,
      "hugging face": 13784,
      "possible use": 25218,
      "models plm": 21818,
      "tasks despite": 32294,
      "argue current": 2490,
      "critical aspect": 6382,
      "modeling human": 20895,
      "human intelligence": 13827,
      "tasks longstanding": 32409,
      "longstanding challenge": 19214,
      "field ai": 11133,
      "crosslingual summarization": 6419,
      "document summary": 8508,
      "open problem": 23415,
      "attention field": 2717,
      "plms gpt2": 25047,
      "models examine": 21222,
      "trained large": 33405,
      "underlying structure": 34163,
      "lms text": 19116,
      "observed model": 23241,
      "model behaviors": 20392,
      "using set": 34909,
      "american english": 1879,
      "consistency large": 5734,
      "does appear": 8522,
      "biases training": 3687,
      "data finetuning": 6715,
      "finetuning t5": 11541,
      "sensitive spelling": 29518,
      "gpt2 similarly": 12951,
      "need largescale": 22635,
      "largescale highquality": 17356,
      "text datasets": 32843,
      "data creation": 6664,
      "dataset spanning": 7039,
      "languages used": 16919,
      "multilingual bloom": 22296,
      "bloom language": 3785,
      "model release": 20750,
      "monolingual multilingual": 22210,
      "data processing": 6808,
      "stimulate research": 31063,
      "large multilingual": 17241,
      "multilingual corpus": 22302,
      "data pruning": 6817,
      "pruning aims": 26809,
      "samples make": 29084,
      "make contribution": 19460,
      "compared original": 5157,
      "original data": 23703,
      "problem propose": 26004,
      "framework aiming": 11826,
      "aiming achieve": 1656,
      "samples based": 29074,
      "classification semantic": 4609,
      "semantic segmentation": 29472,
      "finetuning tasks": 11544,
      "diffusion model": 8183,
      "llama instruction": 18114,
      "selection methods": 29393,
      "methods code": 20010,
      "humans ai": 13919,
      "ai systems": 1618,
      "increasingly important": 14637,
      "systems chatgpt": 31889,
      "chatgpt bloom": 4459,
      "ai writing": 1623,
      "creating image": 6367,
      "tasks present": 32453,
      "present use": 25561,
      "use ai": 34513,
      "holds potential": 13741,
      "traditional kbqa": 33347,
      "analysis question": 1953,
      "llm family": 18303,
      "powerful large": 25342,
      "supports natural": 31718,
      "language question": 16811,
      "answering using": 2080,
      "knowledge growing": 15863,
      "answering kbqa": 2062,
      "models works": 22138,
      "performance chatgpt": 24541,
      "lack largescale": 15997,
      "largescale comprehensive": 17344,
      "comprehensive testing": 5394,
      "various types": 35157,
      "analyze limitations": 1997,
      "limitations model": 17927,
      "family llms": 10981,
      "llms realworld": 18889,
      "complex question": 5286,
      "datasets include": 7130,
      "english datasets": 9478,
      "datasets multilingual": 7150,
      "multilingual datasets": 22306,
      "datasets total": 7182,
      "total number": 33298,
      "test cases": 32762,
      "addition gpt": 1240,
      "llms dataset": 18505,
      "humans language": 13925,
      "chatgpt vicuna": 4497,
      "internal workings": 15443,
      "black box": 3752,
      "unclear llms": 34125,
      "characteristics language": 4426,
      "12 experiments": 90,
      "respectively models": 28462,
      "different words": 8159,
      "sentences likely": 29556,
      "model preferred": 20718,
      "use context": 34520,
      "demonstrate llms": 7472,
      "lesser extent": 17721,
      "extraction given": 10767,
      "extraction aims": 10762,
      "image quality": 14073,
      "formulate task": 11772,
      "task extract": 32122,
      "aspects directly": 2574,
      "relations text": 27871,
      "text paper": 32914,
      "taskoriented dialogs": 32220,
      "systems google": 31900,
      "academic research": 788,
      "research area": 28293,
      "limited lack": 17954,
      "lack datasets": 15982,
      "wide array": 35547,
      "conversations introduce": 6110,
      "public dataset": 26834,
      "conversations humans": 6109,
      "diverse array": 8412,
      "conversational parsing": 6101,
      "dataset provides": 7026,
      "context users": 5925,
      "baselines demonstrate": 3264,
      "demonstrate conversational": 7444,
      "phenomenon present": 24925,
      "challenging model": 4387,
      "grammatical error": 13206,
      "error detection": 9713,
      "detection using": 7886,
      "method detecting": 19901,
      "grammatical errors": 13209,
      "using small": 34915,
      "primarily designed": 25909,
      "designed translation": 7747,
      "task extensive": 32121,
      "optimal performance": 23615,
      "performance final": 24599,
      "presents detailed": 25578,
      "detailed analysis": 7834,
      "challenges adapting": 4334,
      "translation model": 33834,
      "demonstrating potential": 7585,
      "models detecting": 21150,
      "exploring impact": 10616,
      "instruction data": 15143,
      "data scaling": 6848,
      "scaling large": 29167,
      "models empirical": 21192,
      "study realworld": 31389,
      "realworld use": 27351,
      "success chatgpt": 31508,
      "key factor": 15765,
      "remarkable results": 28055,
      "significantly enhances": 30049,
      "enhances models": 9549,
      "makes models": 19493,
      "generated results": 12386,
      "results consistent": 28584,
      "current research": 6527,
      "different amounts": 8043,
      "amounts instruction": 1885,
      "explore performance": 10595,
      "performance large": 24644,
      "based instruction": 3181,
      "tuning different": 33973,
      "different scales": 8135,
      "evaluation dataset": 9936,
      "dataset consisting": 6959,
      "results merely": 28642,
      "continuous improvement": 5997,
      "tasks openended": 32435,
      "math code": 19667,
      "potential future": 25255,
      "selecting highquality": 29387,
      "base models": 3128,
      "training methods": 33565,
      "tasks release": 32479,
      "model checkpoints": 20417,
      "attention placed": 2734,
      "llms downstream": 18548,
      "despite importance": 7787,
      "research space": 28362,
      "corpora using": 6172,
      "highly efficient": 13663,
      "compression rate": 5424,
      "opt 175b": 23587,
      "provides framework": 26754,
      "analysis current": 1916,
      "current future": 6491,
      "output llms": 23873,
      "llms koala": 18739,
      "public use": 26844,
      "llms revolutionized": 18922,
      "revolutionized natural": 28848,
      "impressive capabilities": 14231,
      "capabilities various": 4080,
      "hallucinations model": 13396,
      "false information": 10958,
      "information responses": 14906,
      "llm performance": 18348,
      "performance specific": 24763,
      "specific knowledge": 30699,
      "evaluated based": 9873,
      "based question": 3214,
      "question answer": 27038,
      "answer qa": 2049,
      "dataset covers": 6965,
      "entire field": 9625,
      "model improvement": 20573,
      "likely occur": 17902,
      "assess llms": 2596,
      "technique enables": 32621,
      "enables users": 9311,
      "evaluate llms": 9845,
      "llms performance": 18850,
      "provide detailed": 26693,
      "detailed insights": 7840,
      "insights llms": 15079,
      "llms knowledge": 18734,
      "capabilities different": 4011,
      "datasets llm": 7144,
      "llm responses": 18363,
      "detailed comparison": 7835,
      "comparison multiple": 5197,
      "multiple llms": 22399,
      "conduct comparative": 5587,
      "comparative analysis": 5096,
      "stateoftheart llms": 30945,
      "gpt3 chatgpt": 12987,
      "tested multiple": 32797,
      "feature engineering": 11022,
      "engineering approaches": 9463,
      "feature space": 11028,
      "automated machine": 2865,
      "learning automl": 17543,
      "method utilizes": 19984,
      "linear regression": 17994,
      "llms gpt4": 18654,
      "gptj llama": 13148,
      "llama falcon": 18098,
      "remains important": 27997,
      "important task": 14213,
      "performance increase": 24630,
      "models era": 21212,
      "family parameterefficient": 10983,
      "models success": 22024,
      "like gpt4": 17873,
      "gpt4 chatgpt": 13062,
      "led development": 17684,
      "llms taskspecific": 18997,
      "taskspecific data": 32560,
      "various finetuning": 35095,
      "finetuning peft": 11475,
      "requires finetuning": 28253,
      "llms achieving": 18423,
      "comparable better": 5074,
      "peft methods": 24437,
      "methods llms": 20064,
      "framework integrates": 11870,
      "adapters llms": 1205,
      "llms different": 18539,
      "framework includes": 11866,
      "llms llama": 18767,
      "llama bloom": 18081,
      "promptbased learning": 26361,
      "methods conduct": 20011,
      "evaluate effectiveness": 9828,
      "different reasoning": 8131,
      "arithmetic reasoning": 2504,
      "reasoning commonsense": 27392,
      "demonstrate using": 7511,
      "llms 7b": 18402,
      "parameters yields": 24295,
      "comparable cases": 5076,
      "cases superior": 4207,
      "performance powerful": 24717,
      "powerful llms": 25346,
      "llms 175b": 18401,
      "zeroshot inference": 35980,
      "inference reasoning": 14803,
      "ai community": 1605,
      "generation ai": 12454,
      "combining language": 4963,
      "specific tasks": 30719,
      "like image": 17876,
      "image captioning": 14057,
      "image models": 14072,
      "llm use": 18376,
      "enables better": 9294,
      "serve input": 29648,
      "open ai": 23383,
      "gpt4 demonstrate": 13069,
      "generating novel": 12438,
      "complex constraints": 5268,
      "multimodal models": 22363,
      "models currently": 21119,
      "currently lack": 6542,
      "format task": 11755,
      "task recently": 32186,
      "recently language": 27605,
      "similar problems": 30115,
      "offers enhanced": 23309,
      "enhanced capabilities": 9532,
      "ways work": 35454,
      "models tuned": 22086,
      "feedback large": 11064,
      "chatgpt exhibited": 4465,
      "exhibited remarkable": 10236,
      "remarkable abilities": 28025,
      "language processingnlp": 16804,
      "chat models": 4445,
      "models accessible": 20940,
      "accessible restricted": 834,
      "barriers new": 3104,
      "new research": 22839,
      "field propose": 11151,
      "framework enhance": 11847,
      "based opensource": 3206,
      "opensource llms": 23519,
      "feedback data": 11058,
      "data specifically": 6878,
      "translation data": 33823,
      "translation process": 33846,
      "propose instruction": 26520,
      "including translation": 14524,
      "translation instruction": 33827,
      "instruction contrastive": 15142,
      "instruction experiments": 15153,
      "improves translation": 14397,
      "translation performance": 33844,
      "vanilla llms": 35030,
      "humans demonstrate": 13921,
      "potential automatic": 25243,
      "evaluation tools": 10022,
      "quality information": 26969,
      "human annotation": 13790,
      "pile dataset": 24953,
      "accuracy given": 884,
      "compute budget": 5493,
      "powerlaw scaling": 25356,
      "stateoftheart training": 31001,
      "model scaling": 20767,
      "improving accuracy": 14403,
      "code making": 4778,
      "trained fixed": 33399,
      "available huggingface": 2981,
      "study evaluates": 31327,
      "potential large": 25266,
      "feature norms": 11027,
      "critical tool": 6396,
      "tool evaluating": 33258,
      "cognitive science": 4881,
      "expressed human": 10640,
      "suggest llms": 31574,
      "llms greatly": 18665,
      "methods semantic": 20092,
      "humans machines": 13928,
      "information extraction": 14867,
      "tasks uie": 32536,
      "study various": 31409,
      "extensively utilized": 10720,
      "syntactic knowledge": 31818,
      "better generation": 3606,
      "generation decoding": 12484,
      "finally introduce": 11196,
      "introduce taskoriented": 15536,
      "benchmarks tasks": 3474,
      "tasks shows": 32501,
      "shows significant": 29936,
      "significant improvements": 29991,
      "indepth analyses": 14673,
      "identifying source": 14024,
      "source codes": 30555,
      "educational questions": 8844,
      "questions generated": 27111,
      "generated large": 12367,
      "potential transform": 25303,
      "teachers students": 32588,
      "quality diverse": 26953,
      "diverse question": 8451,
      "generation dramatically": 12489,
      "dramatically reduce": 8719,
      "improve quality": 14290,
      "educational content": 8840,
      "content recent": 5871,
      "real teachers": 27312,
      "classroom setting": 4634,
      "conduct human": 5610,
      "showing promise": 29862,
      "widespread use": 35597,
      "text simplification": 32941,
      "architecture text": 2453,
      "simplification ts": 30174,
      "process generating": 26062,
      "sentences given": 29555,
      "piece text": 24949,
      "text aim": 32815,
      "given text": 12776,
      "english sentences": 9490,
      "proposed architecture": 26593,
      "sentence transformers": 29546,
      "using metrics": 34841,
      "models terms": 22053,
      "terms perplexity": 32749,
      "contributions paper": 6042,
      "baseline research": 3257,
      "results stateoftheart": 28689,
      "better instruction": 3608,
      "instruction following": 15161,
      "following language": 11694,
      "models chinese": 21054,
      "investigating impact": 15609,
      "impact training": 14140,
      "recently significant": 27626,
      "efforts directed": 9089,
      "models capabilities": 21038,
      "conversational models": 6100,
      "models remains": 21926,
      "evaluations models": 10034,
      "performance study": 24770,
      "study examine": 31328,
      "influence training": 14840,
      "quantity quality": 27002,
      "performance analysis": 24523,
      "analysis grounded": 1924,
      "publicly accessible": 26849,
      "highquality instruction": 13690,
      "instruction datasets": 15152,
      "datasets chinese": 7073,
      "chinese multiturn": 4549,
      "multiturn conversations": 22462,
      "using evaluation": 34773,
      "evaluation set": 10008,
      "manual evaluations": 19559,
      "offering valuable": 23303,
      "valuable insights": 35012,
      "opensource chat": 23486,
      "models furthermore": 21298,
      "furthermore enhance": 11998,
      "performance training": 24788,
      "inference efficiency": 14772,
      "llama model": 18128,
      "performance proprietary": 24726,
      "proprietary language": 26637,
      "secondary pretraining": 29331,
      "make model": 19472,
      "comparative study": 5101,
      "finetuning chinese": 11380,
      "following large": 11695,
      "models crucial": 21116,
      "area research": 2483,
      "research field": 28312,
      "parameterefficient tuning": 24216,
      "tuning techniques": 34017,
      "techniques lora": 32650,
      "encouraging results": 9404,
      "fullparameter finetuning": 11949,
      "terms training": 32752,
      "tuning methods": 33999,
      "methods utilizing": 20109,
      "utilizing llama": 34974,
      "llama base": 18078,
      "model experimental": 20500,
      "foundational model": 11808,
      "parameter quantity": 24195,
      "important factors": 14201,
      "provide inspiration": 26711,
      "especially field": 9735,
      "field chinese": 11135,
      "help researchers": 13513,
      "cost model": 6250,
      "negative impact": 22660,
      "blooms taxonomy": 3794,
      "generative text": 12709,
      "ai tools": 1622,
      "concerns regarding": 5546,
      "regarding potential": 27811,
      "concerns paper": 5542,
      "approach aims": 2234,
      "aims identify": 1667,
      "best set": 3578,
      "approach evaluated": 2273,
      "study uses": 31406,
      "data structures": 6883,
      "optimization algorithm": 23622,
      "questions chatgpt": 27101,
      "step forward": 31046,
      "create effective": 6349,
      "efficient effective": 9031,
      "effective text": 8901,
      "llama alpaca": 18073,
      "chatgpt gpt4": 4478,
      "transformed natural": 33699,
      "research shown": 28360,
      "artificial general": 2530,
      "general intelligence": 12168,
      "intelligence agi": 15351,
      "high costs": 13562,
      "costs associated": 6267,
      "associated training": 2652,
      "training deploying": 33499,
      "llms present": 18861,
      "models llama": 21464,
      "predominantly focus": 25462,
      "focus english": 11645,
      "capabilities understanding": 4075,
      "understanding generating": 34224,
      "ability follow": 666,
      "instructions achieve": 15242,
      "tokens improving": 33232,
      "encoding efficiency": 9383,
      "semantic understanding": 29481,
      "pretraining using": 25851,
      "chinese data": 4543,
      "data finetune": 6711,
      "finetune model": 11295,
      "model chinese": 20418,
      "significantly enhancing": 30053,
      "enhancing models": 9570,
      "models ability": 20932,
      "ability comprehend": 659,
      "execute instructions": 10193,
      "results indicate": 28632,
      "newly proposed": 22872,
      "proficiency understanding": 26182,
      "additionally results": 1302,
      "yield competitive": 35909,
      "models times": 22060,
      "open research": 23418,
      "llama series": 18142,
      "llama2 series": 18192,
      "pretraining text": 25846,
      "generalization capabilities": 12208,
      "various downstream": 35089,
      "tasks diverse": 32299,
      "diverse datasets": 8424,
      "datasets large": 7138,
      "corpus containing": 6176,
      "corpus curate": 6178,
      "perform simple": 24503,
      "data filtering": 6709,
      "filtering process": 11173,
      "space using": 30585,
      "use pretrain": 34559,
      "yielding performance": 35919,
      "performance drop": 24577,
      "benchmarks compared": 3433,
      "pretrained checkpoints": 25636,
      "effective instruction": 8876,
      "instructions instruction": 15253,
      "generalize better": 12232,
      "better follow": 3603,
      "follow user": 11680,
      "user intent": 34656,
      "data costly": 6661,
      "examples llms": 10134,
      "llms introduce": 18727,
      "instructions generate": 15252,
      "instructions llms": 15262,
      "select diverse": 29376,
      "set humanwritten": 29689,
      "llms approach": 18436,
      "approach provides": 2330,
      "instructiontuning dataset": 15303,
      "dataset natural": 7016,
      "long text": 19185,
      "outperform 10x": 23762,
      "10x larger": 66,
      "models instruction": 21400,
      "tuning tasks": 34016,
      "longform question": 19205,
      "outperform prior": 23784,
      "instructiontuned models": 15299,
      "models flant5": 21286,
      "improve language": 14270,
      "multilingual instructions": 22309,
      "instructions demonstrate": 15249,
      "news generation": 22882,
      "generation publicly": 12583,
      "release data": 27905,
      "data models": 6776,
      "learning code": 17552,
      "end extract": 9411,
      "applications built": 2143,
      "test potential": 32780,
      "code demonstrate": 4748,
      "critical machine": 6389,
      "draws attention": 8731,
      "learning systems": 17660,
      "extension works": 10672,
      "applications large": 2159,
      "openended questions": 23461,
      "multiplechoice questions": 22431,
      "review answers": 28828,
      "task timeconsuming": 32200,
      "llm paper": 18344,
      "paper analyze": 24011,
      "llms gpt3": 18648,
      "gpt3 bloom": 12982,
      "used zero": 34637,
      "compared performance": 5159,
      "performance results": 24742,
      "llms perform": 18847,
      "perform worse": 24510,
      "questions answers": 27096,
      "faces challenges": 10829,
      "lowrank adaptation": 19300,
      "adaptation lora": 1184,
      "abilities paper": 635,
      "stanford alpaca": 30892,
      "alpaca dataset": 1826,
      "13b 27b": 118,
      "models benchmark": 21012,
      "benchmark models": 3399,
      "models multiple": 21742,
      "multiple ways": 22427,
      "gpt4 judge": 13086,
      "programming tasks": 26203,
      "performant models": 24825,
      "little 40": 18048,
      "zeroshot slot": 35996,
      "slot filling": 30329,
      "taskoriented dialog": 32218,
      "systems work": 31930,
      "lowresource language": 19311,
      "al 2023": 1689,
      "using wide": 34940,
      "range models": 27198,
      "models settings": 21966,
      "settings given": 29737,
      "given recent": 12766,
      "recent success": 27559,
      "test generalization": 32768,
      "capability recent": 4101,
      "recent encoderdecoder": 27519,
      "encoderdecoder model": 9368,
      "model mt0": 20647,
      "al 2022": 1688,
      "languages intentionally": 16879,
      "intentionally seen": 15376,
      "baseline large": 3250,
      "realm computational": 27328,
      "social science": 30431,
      "aim establish": 1640,
      "synthetically generated": 31864,
      "generated data": 12349,
      "data gpt4": 6725,
      "gpt4 llama2": 13091,
      "tasks varying": 32548,
      "varying complexity": 35170,
      "additionally examine": 1283,
      "examine impact": 10101,
      "data sizes": 6869,
      "findings reveal": 11248,
      "reveal models": 28804,
      "data consistently": 6654,
      "exhibit superior": 10231,
      "tasks furthermore": 32343,
      "leverage gpt4": 17748,
      "fall short": 10950,
      "short compared": 29811,
      "compared specialized": 5169,
      "moderately sized": 22152,
      "peft techniques": 24438,
      "techniques llms": 32649,
      "currently popular": 6544,
      "provide comprehensive": 26690,
      "benchmark various": 3419,
      "representative llm": 28183,
      "llm flant5": 18308,
      "flant5 model": 11599,
      "model evaluate": 20494,
      "performance different": 24567,
      "different data": 8064,
      "data scales": 6847,
      "classification generation": 4598,
      "generation datasets": 12483,
      "datasets based": 7068,
      "provide framework": 26703,
      "optimal finetuning": 23612,
      "finetuning techniques": 11546,
      "given task": 12774,
      "data availability": 6617,
      "techniques applied": 32627,
      "parameters maintaining": 24272,
      "maintaining improving": 19425,
      "outperforming larger": 23798,
      "deploying large": 7639,
      "llms challenging": 18462,
      "train smaller": 33375,
      "taskspecific models": 32566,
      "finetuning distillation": 11390,
      "amounts training": 1890,
      "data achieve": 6588,
      "achieve comparable": 952,
      "llms achieves": 18422,
      "needed finetuning": 22647,
      "distillation method": 8344,
      "supervision training": 31699,
      "training small": 33616,
      "multitask framework": 22447,
      "framework present": 11887,
      "compared finetuning": 5134,
      "distillation mechanism": 8343,
      "llms achieve": 18410,
      "substantially smaller": 31487,
      "reduce model": 27719,
      "size data": 30243,
      "palm model": 23992,
      "available data": 2970,
      "data benchmark": 6622,
      "standard finetuning": 30876,
      "using 100": 34722,
      "dataset release": 7032,
      "entity tracking": 9654,
      "discourse entities": 8264,
      "given english": 12748,
      "gpt3 gpt35": 12996,
      "performance degrades": 24564,
      "evaluated different": 9875,
      "different set": 8137,
      "entities training": 9642,
      "training longer": 33557,
      "taken results": 32025,
      "suggest language": 31571,
      "models learn": 21439,
      "nucleus sampling": 23136,
      "sampling language": 29091,
      "text based": 32821,
      "set words": 29716,
      "work assess": 35671,
      "various linguistic": 35111,
      "conformal prediction": 5669,
      "prediction calibration": 25422,
      "calibration procedure": 3980,
      "prediction sets": 25435,
      "opt models": 23600,
      "inverse scaling": 15568,
      "intelligence ai": 15352,
      "image generators": 14068,
      "complex art": 5266,
      "text generators": 32889,
      "including chatgpt": 14464,
      "allow users": 1802,
      "generate code": 12266,
      "current state": 6530,
      "addresses main": 1366,
      "approach taken": 2345,
      "ai article": 1598,
      "models despite": 21144,
      "performs worse": 24858,
      "tuning simple": 34014,
      "method significantly": 19973,
      "soft prompt": 30448,
      "prompt embeddings": 26321,
      "residual connection": 28389,
      "notably method": 23030,
      "points improvement": 25075,
      "allows reduce": 1816,
      "prompt length": 26332,
      "hurting performance": 13953,
      "performance addition": 24517,
      "addition approach": 1236,
      "rate prompt": 27267,
      "effective fewshot": 8869,
      "era large": 9697,
      "semantic relationships": 29467,
      "text standard": 32949,
      "training modules": 33570,
      "conditioned input": 5576,
      "using larger": 34815,
      "work evaluating": 35700,
      "evaluating performance": 9909,
      "varying levels": 35176,
      "generative approaches": 12650,
      "sota performance": 30538,
      "supervised models": 31688,
      "chainofthought cot": 4297,
      "generated gpt3": 12360,
      "results release": 28670,
      "release model": 27910,
      "model new": 20657,
      "new baseline": 22779,
      "contrastive objective": 6023,
      "text embeddings": 32850,
      "useful features": 34639,
      "semantic search": 29471,
      "produce semantically": 26153,
      "second finetune": 29321,
      "adapter lora": 1201,
      "data scarcity": 6850,
      "quality learned": 26972,
      "learned embeddings": 17519,
      "proportional number": 26484,
      "unlabeled training": 34382,
      "data parameter": 6794,
      "previous solution": 25877,
      "achieve significant": 982,
      "models ptlms": 21876,
      "remarkable improvement": 28043,
      "emergence new": 9174,
      "new capabilities": 22783,
      "pretraining dataset": 25792,
      "training times": 33635,
      "environmental impact": 9667,
      "significant efforts": 29980,
      "training efficient": 33508,
      "training pipelines": 33588,
      "attention paid": 2729,
      "data key": 6744,
      "question ask": 27062,
      "data maintaining": 6763,
      "building recent": 3926,
      "subset selection": 31452,
      "corpora demonstrate": 6165,
      "framework applied": 11830,
      "efficiently train": 9074,
      "train multiple": 33369,
      "bert biobert": 3499,
      "data perform": 6795,
      "intent classification": 15372,
      "scale thousands": 29148,
      "coldstart problem": 4901,
      "space paper": 30580,
      "approaches lowresource": 2383,
      "adaptation data": 1176,
      "classification using": 4619,
      "descriptions large": 7688,
      "models results": 21941,
      "results approaches": 28571,
      "approaches effective": 2373,
      "different degrees": 8069,
      "yields best": 35921,
      "best performance": 3568,
      "performance just": 24637,
      "zeroshot method": 35985,
      "transforming natural": 33804,
      "natural languages": 22577,
      "models temporal": 22048,
      "systems engineering": 31895,
      "engineering applications": 9462,
      "language nl": 16764,
      "underexplored lack": 34141,
      "lack dataset": 15981,
      "model different": 20470,
      "different application": 8044,
      "application domains": 2128,
      "llms multiple": 18809,
      "multiple stages": 22421,
      "contributions twofold": 6044,
      "create dataset": 6348,
      "combining llms": 4964,
      "llms human": 18678,
      "characterizes common": 4432,
      "domains application": 8611,
      "application llms": 2134,
      "dataset creation": 6967,
      "varied domains": 35051,
      "recognition task": 27640,
      "finetuning specific": 11531,
      "specific domain": 30688,
      "domain finetuning": 8564,
      "finetuning model": 11454,
      "achieves higher": 1046,
      "higher accuracy": 13594,
      "accuracy 95": 861,
      "using 10": 34721,
      "evaluation platform": 9989,
      "interaction user": 15386,
      "user interface": 34659,
      "human interaction": 13829,
      "digital world": 8191,
      "facilitating efficient": 10853,
      "task completion": 32096,
      "researchers exploring": 28377,
      "exploring potential": 10621,
      "programming language": 26199,
      "language interfaces": 16100,
      "graphical user": 13234,
      "user interfaces": 34660,
      "interfaces guis": 15417,
      "limited capabilities": 17941,
      "models traditional": 22062,
      "single step": 30223,
      "llms exhibited": 18582,
      "reasoning planning": 27435,
      "planning abilities": 25000,
      "complex environments": 5275,
      "remains underexplored": 28018,
      "benchmark covering": 3366,
      "interaction capabilities": 15384,
      "comprehensive evaluations": 5375,
      "llm agents": 18266,
      "gpt llama": 12853,
      "challenges llms": 4359,
      "identifying causal": 14020,
      "generalpurpose language": 12248,
      "ai safety": 1616,
      "underlying model": 34160,
      "generalize unseen": 12234,
      "unseen inputs": 34437,
      "finetuned specific": 11353,
      "present paper": 25550,
      "search steps": 29313,
      "learned parameters": 17520,
      "parameters approach": 24225,
      "alpaca model": 1832,
      "7b parameters": 542,
      "numerical reasoning": 23179,
      "reasoning problem": 27437,
      "alignment neural": 1775,
      "neural representations": 22757,
      "widely deployed": 35572,
      "deployed language": 7633,
      "models tool": 22061,
      "larger llms": 17326,
      "released publicly": 27928,
      "leveraging large": 17783,
      "prediction large": 25426,
      "llms produce": 18871,
      "produce humanlike": 26149,
      "humanlike responses": 13908,
      "social sciences": 30433,
      "integrate llms": 15322,
      "llms social": 18958,
      "questions asked": 27098,
      "asked develop": 2556,
      "questions derived": 27106,
      "text latent": 32907,
      "contexts different": 5937,
      "finetuning llms": 11448,
      "using general": 34783,
      "ethical concerns": 9804,
      "prediction study": 25436,
      "study demonstrates": 31316,
      "enhance capabilities": 9505,
      "capabilities llms": 4040,
      "alignment llms": 1768,
      "enhancing large": 9562,
      "longterm memory": 19217,
      "advancements large": 1465,
      "intelligence systems": 15361,
      "increasingly evident": 14636,
      "tailored llms": 32017,
      "enables models": 9305,
      "synthesizing information": 31844,
      "updating mechanism": 34478,
      "closedsource models": 4688,
      "chatgpt opensource": 4491,
      "llmbased chatbot": 18385,
      "chatbot named": 4450,
      "realworld user": 27353,
      "covering wide": 6328,
      "topics results": 33292,
      "results analysis": 28569,
      "exhibits strong": 10254,
      "strong capability": 31166,
      "understand user": 34202,
      "generate factually": 12276,
      "correct answers": 6192,
      "answers existing": 2083,
      "existing llms": 10290,
      "llms generate": 18634,
      "responses different": 28490,
      "different prompts": 8130,
      "prompts paper": 26434,
      "study problem": 31379,
      "knowledge contained": 15826,
      "facts propose": 10875,
      "knowledge llms": 15880,
      "llms main": 18789,
      "main idea": 19398,
      "llm generating": 18316,
      "text corresponding": 32838,
      "comprehensive set": 5390,
      "method evaluate": 19916,
      "20 llms": 212,
      "llms various": 19034,
      "various sizes": 35144,
      "including llama": 14498,
      "results strong": 28691,
      "strong correlation": 31167,
      "results human": 28623,
      "human assessment": 13794,
      "assessment llms": 2619,
      "llms results": 18918,
      "backbone architecture": 3053,
      "scaling law": 29171,
      "instructionfollowing data": 15226,
      "models capability": 21039,
      "capability generate": 4089,
      "correct text": 6194,
      "realworld settings": 27348,
      "settings work": 29745,
      "task approach": 32078,
      "generation generate": 12509,
      "indomain dataset": 14724,
      "dataset using": 7053,
      "prompt training": 26351,
      "datasets improve": 7129,
      "able generalize": 744,
      "seen training": 29366,
      "alignment large": 1765,
      "unsupervised pretraining": 34457,
      "representations large": 28162,
      "learning better": 17546,
      "better align": 3587,
      "tasks user": 32543,
      "user preferences": 34664,
      "65b parameter": 477,
      "llama language": 18115,
      "finetuned standard": 11354,
      "learning human": 17575,
      "human preference": 13853,
      "remarkably strong": 28067,
      "performance learning": 24653,
      "follow specific": 11678,
      "specific response": 30713,
      "handful examples": 13405,
      "data including": 6734,
      "complex queries": 5285,
      "controlled human": 6064,
      "human study": 13866,
      "human feedback": 13818,
      "suggest knowledge": 31570,
      "models learned": 21441,
      "learned pretraining": 17521,
      "pretraining limited": 25815,
      "tuning data": 33970,
      "data necessary": 6782,
      "models produce": 21856,
      "produce high": 26146,
      "instructions prompting": 15264,
      "instructions recently": 15268,
      "emerged popular": 9162,
      "harnessing capabilities": 13459,
      "capabilities large": 4031,
      "explore prompting": 10599,
      "models manually": 21700,
      "tasks spanning": 32511,
      "tasks sourced": 32510,
      "language study": 16826,
      "performance llm": 24662,
      "llm families": 18301,
      "families bloom": 10965,
      "instructions leads": 15260,
      "results average": 28574,
      "average increase": 3021,
      "f1 scores": 10818,
      "tasks improvement": 32360,
      "improvement relative": 14346,
      "rougel scores": 28983,
      "tasks include": 32362,
      "detailed ablation": 7831,
      "encoded pseudocode": 9344,
      "improvement performance": 14344,
      "performance best": 24534,
      "work demonstrate": 35690,
      "lms nlp": 19099,
      "pretraining large": 25810,
      "large corpora": 16937,
      "corpora text": 6170,
      "models acquire": 20955,
      "achieve remarkable": 981,
      "acquire knowledge": 1119,
      "question paper": 27073,
      "ability models": 705,
      "systematically create": 31878,
      "evaluation data": 9935,
      "data data": 6669,
      "questions based": 27099,
      "evaluations multiple": 10035,
      "flan t5": 11588,
      "performance gap": 24607,
      "revealing interesting": 28810,
      "interesting findings": 15408,
      "research developing": 28306,
      "developing robust": 7947,
      "attention layers": 2722,
      "study discover": 31319,
      "networks used": 22717,
      "based finding": 3159,
      "finding propose": 11226,
      "summarization experiments": 31611,
      "best method": 3562,
      "based t5small": 3228,
      "xsum dataset": 35882,
      "dataset improves": 7000,
      "different architectures": 8049,
      "different popular": 8121,
      "popular autoregressive": 25113,
      "models alms": 20976,
      "important application": 14196,
      "successfully used": 31545,
      "cnn lstm": 4712,
      "transformer networks": 33736,
      "different training": 8152,
      "methods investigate": 20054,
      "investigate capabilities": 15578,
      "recognition using": 27643,
      "classification higher": 4599,
      "model robustness": 20763,
      "emergence generative": 9168,
      "generative large": 12666,
      "llms raises": 18885,
      "raises question": 27169,
      "including ones": 14508,
      "evaluation tasks": 10019,
      "human workers": 13877,
      "paraphrase generation": 24301,
      "apply data": 2205,
      "data collection": 6642,
      "similar scale": 30117,
      "seed data": 29352,
      "using chatgpt": 34749,
      "lead robust": 17469,
      "models response": 21938,
      "llm inference": 18321,
      "revolutionized field": 28844,
      "inference process": 14800,
      "process llms": 26072,
      "llms comes": 18472,
      "comes significant": 4971,
      "costs paper": 6272,
      "propose efficient": 26507,
      "efficient llm": 9044,
      "pipeline harnesses": 24968,
      "harnesses power": 13457,
      "power llms": 25324,
      "potential llms": 25274,
      "llms accurately": 18409,
      "information introduce": 14874,
      "introduce efficient": 15504,
      "evaluate approach": 9819,
      "approach realworld": 2331,
      "llamabased model": 18237,
      "inference throughput": 14816,
      "inference acceleration": 14759,
      "techniques making": 32651,
      "making valuable": 19518,
      "quantization llm": 27008,
      "multitask instruction": 22448,
      "tuning llama": 33992,
      "specific scenarios": 30714,
      "preliminary study": 25491,
      "writing assistance": 35850,
      "proprietary large": 26639,
      "garnered significant": 12122,
      "significant attention": 29960,
      "attention exceptional": 2716,
      "exceptional capabilities": 10167,
      "handling diverse": 13415,
      "range tasks": 27212,
      "studies demonstrate": 31264,
      "foundational models": 11809,
      "llama display": 18092,
      "display remarkable": 8315,
      "remarkable proficiency": 28053,
      "tasks finetuned": 32335,
      "finetuned using": 11359,
      "data work": 6915,
      "practical problem": 25369,
      "primary focus": 25921,
      "particular tasks": 24343,
      "tasks generalpurpose": 32346,
      "explore llms": 10591,
      "llm specifically": 18370,
      "tuning experimental": 33977,
      "finetuning llama": 11439,
      "data significantly": 6864,
      "conduct experiments": 5602,
      "offer insights": 23289,
      "insights future": 15074,
      "effectively finetuning": 8918,
      "employing llms": 9260,
      "taking account": 32036,
      "evaluating large": 9899,
      "models spoken": 22004,
      "spoken language": 30815,
      "demonstrated strong": 7549,
      "capabilities particularly": 4060,
      "tasks prompting": 32461,
      "assess impact": 2595,
      "understanding slu": 34270,
      "opt different": 23590,
      "multiple benchmarks": 22380,
      "emergent ability": 9183,
      "models reach": 21895,
      "classification accuracy": 4591,
      "models zero": 22140,
      "various languages": 35109,
      "contrast results": 6017,
      "results smaller": 28686,
      "fall far": 10949,
      "error cases": 9710,
      "annotation scheme": 2026,
      "responses chatgpt": 28487,
      "textual models": 33034,
      "encyclopedic knowledge": 9407,
      "assess ability": 2587,
      "ability foundation": 668,
      "range linguistic": 27196,
      "dataset contains": 6963,
      "paired counterfactuals": 23973,
      "models multilingual": 21740,
      "24 models": 263,
      "metas llama": 19863,
      "llama achieves": 18071,
      "errors reveals": 9727,
      "reveals significant": 28818,
      "languages english": 16871,
      "overall findings": 23906,
      "models far": 21263,
      "knowledge largescale": 15875,
      "task planning": 32176,
      "major challenge": 19439,
      "work exploits": 35703,
      "llms directly": 18542,
      "interesting results": 15411,
      "results paper": 28653,
      "shows llms": 29928,
      "llms provide": 18879,
      "world model": 35835,
      "search algorithm": 29307,
      "monte carlo": 22215,
      "carlo tree": 4184,
      "tree search": 33878,
      "model provides": 20736,
      "achieve effective": 961,
      "vastly improving": 35192,
      "llms gpt2": 18647,
      "gpt2 gpt35": 12903,
      "gpt35 wide": 13039,
      "novel tasks": 23114,
      "policy using": 25090,
      "using llm": 34824,
      "llm world": 18380,
      "better using": 3636,
      "finegrained atomic": 11270,
      "form text": 11742,
      "longform text": 19208,
      "mixture supported": 20282,
      "judgments quality": 15728,
      "quality inadequate": 26968,
      "new evaluation": 22799,
      "atomic facts": 2681,
      "extensive human": 10707,
      "stateoftheart commercial": 30925,
      "instructgpt chatgpt": 15137,
      "chatgpt retrievalaugmented": 4495,
      "report new": 28121,
      "new analysis": 22774,
      "demonstrating need": 7584,
      "model error": 20492,
      "finally use": 11205,
      "new set": 22842,
      "set 13": 29670,
      "recent lms": 27532,
      "evaluated humans": 9878,
      "public models": 26840,
      "models vicuna": 22114,
      "alpaca best": 1825,
      "best public": 3576,
      "enhanced crosslingual": 9533,
      "explores potential": 10613,
      "potential leveraging": 25271,
      "llms data": 18503,
      "datasets available": 7067,
      "data extremely": 6707,
      "effectiveness finetuning": 8945,
      "finetuning smaller": 11529,
      "multilingual models": 22320,
      "models mbert": 21707,
      "data generated": 6721,
      "target languages": 32053,
      "accuracy score": 910,
      "score improvement": 29270,
      "improvement best": 14333,
      "furthermore conduct": 11989,
      "coherence generated": 4889,
      "generated examples": 12354,
      "different languages": 8090,
      "languages results": 16914,
      "indicate llms": 14689,
      "coherent text": 4896,
      "text languages": 32902,
      "struggle generate": 31240,
      "text certain": 32823,
      "certain languages": 4275,
      "languages like": 16888,
      "like tamil": 17896,
      "falls short": 10956,
      "generating plausible": 12439,
      "gpt4 exhibit": 13074,
      "logical consistency": 19156,
      "meets llm": 19785,
      "need understand": 22645,
      "understanding reduce": 34267,
      "automatic speech": 2894,
      "recognition asr": 27637,
      "approach focuses": 2285,
      "interactions users": 15392,
      "users history": 34691,
      "history present": 13730,
      "challenges personalized": 4367,
      "rewriting paper": 28869,
      "approach builds": 2245,
      "user feedback": 34649,
      "graph traversal": 13231,
      "retrieval model": 28747,
      "ranking model": 27233,
      "utilization large": 34955,
      "link prediction": 18028,
      "domains specifically": 8640,
      "specifically paper": 30750,
      "7b model": 536,
      "augmented finetuned": 2813,
      "significantly enhanced": 30048,
      "queries compared": 27019,
      "bias large": 3650,
      "cater diverse": 4233,
      "diverse cultural": 8421,
      "despite advancements": 7768,
      "multilingual capabilities": 22297,
      "capabilities models": 4045,
      "models designed": 21143,
      "exhibit bias": 10212,
      "camel novel": 3988,
      "cultural biases": 6459,
      "intrinsic evaluations": 15491,
      "evaluations using": 10041,
      "16 different": 157,
      "lms tasks": 19115,
      "tasks story": 32514,
      "story generation": 31088,
      "ner sentiment": 22678,
      "finally analyze": 11188,
      "lms used": 19121,
      "present systematic": 25558,
      "systematic study": 31875,
      "particular construct": 24335,
      "dataset human": 6999,
      "english french": 9480,
      "data sources": 6875,
      "domain language": 8570,
      "language diversity": 16066,
      "nonenglish language": 22991,
      "llama2 gpt4": 18180,
      "gpt4 supervised": 13121,
      "supervised unsupervised": 31693,
      "prompting settings": 26394,
      "settings experiments": 29733,
      "experiments reveal": 10478,
      "datasets showcasing": 7171,
      "readability metrics": 27295,
      "grade level": 13184,
      "stateoftheart unsupervised": 31005,
      "make data": 19461,
      "hallucination large": 13378,
      "models inference": 21397,
      "tasks large": 32390,
      "capable natural": 4114,
      "like question": 17892,
      "present series": 25552,
      "studies llm": 31274,
      "llama gpt35": 18110,
      "gpt35 palm": 13032,
      "generative llms": 12671,
      "test samples": 32782,
      "memorized data": 19797,
      "similar effect": 30100,
      "data bias": 6626,
      "previous studies": 25885,
      "perform significantly": 24501,
      "significantly worse": 30089,
      "future llm": 12038,
      "llm evaluation": 18297,
      "decoding language": 7274,
      "lms struggle": 19113,
      "pay attention": 24425,
      "context generate": 5891,
      "generate texts": 12332,
      "contain hallucinations": 5827,
      "hallucinations mitigate": 13395,
      "mitigate issue": 20254,
      "issue present": 15660,
      "output distribution": 23865,
      "output probabilities": 23877,
      "used context": 34591,
      "context experiments": 5889,
      "different lm": 8100,
      "families including": 10967,
      "including opt": 14511,
      "opt gpt": 23592,
      "llama flant5": 18102,
      "summarization tasks": 31626,
      "factuality metrics": 10893,
      "metrics furthermore": 20138,
      "particularly effective": 24346,
      "models prior": 21854,
      "prior knowledge": 25936,
      "provided context": 26737,
      "leading substantial": 17487,
      "improvements tasks": 14365,
      "adapting language": 1207,
      "lms powerful": 19100,
      "context window": 5926,
      "expensive computational": 10358,
      "text documents": 32847,
      "propose adapt": 26490,
      "models capable": 21040,
      "compressing long": 5410,
      "long contexts": 19169,
      "model soft": 20802,
      "used language": 34606,
      "opt llama2": 23597,
      "llama2 models": 18188,
      "models sequences": 21965,
      "contexts improve": 5940,
      "task demonstrations": 32107,
      "increasing accuracy": 14617,
      "inference costs": 14771,
      "retrievalaugmented language": 28763,
      "passage reranking": 24379,
      "extracts comprehensive": 10783,
      "entities relations": 9639,
      "seek develop": 29358,
      "llm able": 18260,
      "using instruction": 34800,
      "tuning particular": 34004,
      "tuning dataset": 33971,
      "extensive annotations": 10678,
      "annotations diverse": 2031,
      "strong instructionfollowing": 31175,
      "instructionfollowing capabilities": 15224,
      "capabilities experiments": 4016,
      "outperforms traditional": 23862,
      "methods llm": 20063,
      "capabilities unseen": 4077,
      "unseen instructions": 34438,
      "emerges promising": 9188,
      "solution tackle": 30481,
      "recent advancements": 27489,
      "language speech": 16824,
      "research despite": 28305,
      "despite progress": 7802,
      "models lack": 21413,
      "lack specific": 16002,
      "models tailored": 22045,
      "speech processing": 30785,
      "models gpt35turbo": 21331,
      "gpt35turbo gpt4": 13043,
      "gpt4 bloomz": 13060,
      "distinct tasks": 8369,
      "46 hours": 398,
      "texttospeech tts": 33009,
      "experiments analysis": 10419,
      "measuring performance": 19744,
      "trend observed": 33888,
      "outperformed llms": 23793,
      "llms zeroshot": 19050,
      "computational models": 5472,
      "performance gaps": 24609,
      "findings provide": 11246,
      "provide valuable": 26734,
      "applicability llms": 2123,
      "semantic textual": 29479,
      "textual similarity": 33038,
      "degree similarity": 7390,
      "pair sentences": 23971,
      "broad application": 3884,
      "depending specific": 7621,
      "proposing novel": 26633,
      "task called": 32089,
      "described natural": 7678,
      "large small": 17278,
      "enables finegrained": 9298,
      "model evaluation": 20495,
      "evaluation diverse": 9942,
      "diverse natural": 8441,
      "flant5 gpt4": 11596,
      "spearman correlation": 30652,
      "evaluation semantic": 10007,
      "examples code": 10119,
      "significantly improved": 30058,
      "improved instruction": 14312,
      "lack transparency": 16008,
      "ability utilize": 729,
      "uptodate knowledge": 34490,
      "generation instruction": 12524,
      "following abilities": 11686,
      "search results": 29311,
      "results generated": 28619,
      "results training": 28698,
      "construct new": 5802,
      "set containing": 29679,
      "information response": 14905,
      "finetune llama7b": 11291,
      "llama7b model": 18234,
      "languages model": 16898,
      "model needs": 20655,
      "needs learn": 22652,
      "generate target": 12327,
      "multihop reasoning": 22288,
      "reasoning retrieved": 27449,
      "retrieved passages": 28776,
      "answer experiments": 2044,
      "experiments finetuned": 10442,
      "model strong": 20810,
      "instructionfollowing ability": 15223,
      "performs significantly": 24854,
      "openended question": 23459,
      "answering fact": 2061,
      "unified multilingual": 34333,
      "multilingual benchmark": 22295,
      "automatic text": 2899,
      "research work": 28368,
      "benchmark covers": 3367,
      "pairs benchmark": 23976,
      "developing effective": 7942,
      "effective multilingual": 8888,
      "simplification models": 30173,
      "metrics experiments": 20137,
      "pretrained multilingual": 25732,
      "models reveal": 21944,
      "transfer lowresource": 33679,
      "comparable quality": 5088,
      "validate findings": 34996,
      "proprietary llms": 26645,
      "llms emerging": 18556,
      "model finetune": 20523,
      "stronger model": 31199,
      "chatgpt alpaca": 4457,
      "proprietary models": 26647,
      "capabilities using": 4078,
      "weaker opensource": 35459,
      "model work": 20871,
      "critically analyze": 6399,
      "approach finetune": 2281,
      "output quality": 23878,
      "following instructions": 11692,
      "chatgpt conducting": 4461,
      "automatic evaluations": 2881,
      "base lm": 3123,
      "tasks heavily": 32353,
      "performance discrepancies": 24568,
      "models adept": 20963,
      "gap open": 12098,
      "open closed": 23390,
      "lms current": 19077,
      "current methods": 6514,
      "using capable": 34744,
      "base lms": 3124,
      "models tackle": 22044,
      "better base": 3592,
      "proprietary systems": 26651,
      "highquality dataset": 13685,
      "tasks unlike": 32539,
      "unlike prior": 34400,
      "prior works": 25943,
      "works rely": 35825,
      "extremescale teacher": 10807,
      "teacher model": 32584,
      "model gpt3": 20557,
      "produces highquality": 26162,
      "evaluate method": 9848,
      "method multiple": 19948,
      "generation sentence": 12600,
      "summarization model": 31620,
      "consistently outperforms": 5755,
      "baselines including": 3267,
      "distilled chatgpt": 8350,
      "chatgpt distilled": 4464,
      "distilled dataset": 8353,
      "exhibits higher": 10247,
      "times larger": 33164,
      "small finetuned": 30341,
      "models methods": 21721,
      "whitebox access": 35542,
      "access model": 822,
      "recent trend": 27565,
      "highest quality": 13621,
      "quality models": 26973,
      "weights available": 35503,
      "lightweight method": 17841,
      "new domains": 22795,
      "domains tasks": 8642,
      "intermediate activations": 15424,
      "approach finetunes": 2282,
      "finetunes small": 11365,
      "learned small": 17523,
      "small validation": 30372,
      "validate approach": 34992,
      "large lm": 17225,
      "task machine": 32154,
      "performance cases": 24539,
      "models partially": 21804,
      "interpretation large": 15463,
      "large body": 16933,
      "body literature": 3805,
      "llms acquire": 18424,
      "linguistic representations": 18021,
      "little known": 18051,
      "linguistic biases": 18006,
      "way present": 35446,
      "asking llms": 2559,
      "biases using": 3688,
      "using stimuli": 34921,
      "psycholinguistic experiments": 26821,
      "experiments recent": 10475,
      "studies suggest": 31287,
      "learning icl": 17583,
      "semantic biases": 29448,
      "fails generate": 10914,
      "results provide": 28663,
      "provide evidence": 26697,
      "contemporary llms": 5846,
      "sensitive syntactic": 29519,
      "local context": 19128,
      "semantic patterns": 29464,
      "planning capabilities": 25001,
      "capabilities pretrained": 4061,
      "models wide": 22127,
      "studies ability": 31262,
      "gpt2 empirically": 12886,
      "empirically demonstrate": 9247,
      "demonstrate performance": 7478,
      "finetuned baseline": 11304,
      "capabilities finetuned": 4019,
      "finetuned llm": 11338,
      "train verifier": 33378,
      "dataset generate": 6995,
      "invalid trajectories": 15562,
      "success rate": 31523,
      "domain additionally": 8553,
      "additionally finetuning": 1287,
      "better finetuning": 3602,
      "finetuning base": 11375,
      "base gpt2": 3116,
      "lastly investigate": 17399,
      "investigate role": 15597,
      "sampling temperature": 29099,
      "used control": 34592,
      "data makes": 6764,
      "text gpt2": 32890,
      "gpt4 demonstrated": 13070,
      "astonishing performance": 2668,
      "general public": 12183,
      "ecosystem online": 8815,
      "online text": 23371,
      "text images": 32895,
      "images paper": 14086,
      "paper consider": 24026,
      "llms contribute": 18492,
      "language online": 16765,
      "original content": 23701,
      "content distribution": 5858,
      "effect model": 8854,
      "model collapse": 20428,
      "variational autoencoders": 35046,
      "gaussian mixture": 12132,
      "mixture models": 20281,
      "largescale data": 17345,
      "data collected": 6641,
      "human interactions": 13830,
      "content generated": 5860,
      "data crawled": 6662,
      "dont know": 8664,
      "excel various": 10155,
      "research focuses": 28318,
      "existing knowledge": 10280,
      "vast knowledge": 35186,
      "llms limited": 18766,
      "limited information": 17950,
      "ability understand": 724,
      "understand limitations": 34194,
      "paramount importance": 24299,
      "ability identify": 684,
      "identify unanswerable": 14018,
      "questions introduce": 27115,
      "responses models": 28501,
      "providing novel": 26779,
      "unique dataset": 34359,
      "diverse categories": 8415,
      "counterparts extensive": 6303,
      "extensive analysis": 10677,
      "llms including": 18688,
      "gpt3 instructgpt": 12999,
      "demonstrate incontext": 7464,
      "learning instruction": 17588,
      "tuning enhance": 33975,
      "despite promising": 7803,
      "findings highlight": 11235,
      "considerable gap": 5709,
      "promptbased fewshot": 26359,
      "fewshot finetuning": 11103,
      "finetuning recent": 11502,
      "years significant": 35898,
      "progress developing": 26210,
      "finetuned small": 11352,
      "small datasets": 30339,
      "datasets address": 7061,
      "adaptation approaches": 1174,
      "promptbased tuning": 26365,
      "way especially": 35429,
      "shows adding": 29920,
      "contrastive learning": 6022,
      "promptbased finetuning": 26360,
      "finetuning effective": 11393,
      "helps model": 13527,
      "generate embeddings": 12274,
      "negative examples": 22659,
      "important components": 14199,
      "learning data": 17556,
      "effective data": 8864,
      "especially large": 9739,
      "experiments multiple": 10463,
      "multiple text": 22425,
      "augmentation method": 2802,
      "outperforms methods": 23833,
      "models handle": 21347,
      "word frequency": 35640,
      "prediction head": 25424,
      "reveal biases": 28790,
      "word prediction": 35643,
      "prediction heads": 25425,
      "significant role": 30023,
      "adjustment method": 1394,
      "autoregressive text": 2954,
      "generation scenarios": 12599,
      "scenarios particular": 29213,
      "text quality": 32925,
      "recent efforts": 27516,
      "models retrieval": 21943,
      "data input": 6737,
      "added training": 1230,
      "training test": 33630,
      "computation memory": 5448,
      "memory grows": 19817,
      "data test": 6891,
      "time using": 33146,
      "using standard": 34919,
      "standard training": 30883,
      "build largescale": 3914,
      "largescale distributed": 17350,
      "test input": 32771,
      "performance 20": 24514,
      "gap small": 12110,
      "10 times": 31,
      "quality size": 26978,
      "work establishes": 35698,
      "learning various": 17669,
      "various design": 35086,
      "settings incontext": 29738,
      "bias model": 3657,
      "model particular": 20693,
      "understanding task": 34273,
      "design choices": 7699,
      "mitigating impact": 20265,
      "work define": 35688,
      "analysis demonstrates": 1917,
      "calibration methods": 3979,
      "methods fall": 20035,
      "short addressing": 29810,
      "types biases": 34058,
      "calibration method": 3978,
      "using random": 34896,
      "making predictions": 19515,
      "icl performance": 13977,
      "gptj gpt3": 13147,
      "pretraining methods": 25820,
      "instructions showing": 15272,
      "exploratory study": 10569,
      "study using": 31407,
      "given rapid": 12765,
      "llms study": 18973,
      "models help": 21351,
      "scientific papers": 29253,
      "gpt4 outperforms": 13107,
      "outperforms llms": 23831,
      "bard vicuna": 3100,
      "alpaca llama": 1831,
      "study use": 31405,
      "llms specifically": 18964,
      "specifically gpt4": 30744,
      "gpt4 tasks": 13123,
      "errors construct": 9723,
      "computer science": 5504,
      "check correctness": 4514,
      "errors spanning": 9728,
      "pairs llm": 23981,
      "better paper": 3614,
      "paper generate": 24055,
      "pairs based": 23975,
      "based experiments": 3156,
      "llms promising": 18873,
      "comparing language": 5190,
      "models challenging": 21051,
      "challenging current": 4379,
      "method enables": 19910,
      "topics demonstrate": 33290,
      "similarities differences": 30124,
      "distilroberta gpt2": 8361,
      "web data": 35476,
      "data web": 6912,
      "models commonly": 21066,
      "commonly trained": 5028,
      "trained mixture": 33411,
      "data curated": 6665,
      "curated highquality": 6471,
      "highquality corpora": 13683,
      "curation process": 6476,
      "models broad": 21034,
      "models requiring": 21932,
      "data lead": 6753,
      "powerful models": 25347,
      "significantly outperforming": 30073,
      "outperforming models": 23800,
      "despite extensive": 7776,
      "trillion tokens": 33906,
      "billion tokens": 3721,
      "gpt model": 12856,
      "largescale transformer": 17383,
      "modelling tasks": 20915,
      "models feature": 21265,
      "parameters leading": 24266,
      "prohibitive training": 26237,
      "forward backward": 11782,
      "training resulting": 33601,
      "including language": 14496,
      "understanding text": 34276,
      "model performs": 20709,
      "model proposed": 20733,
      "pretraining transformer": 25849,
      "focused enhancing": 11663,
      "imitation learning": 14110,
      "outputs generated": 23890,
      "large foundation": 16945,
      "outputs small": 23898,
      "small scale": 30365,
      "tend learn": 32706,
      "reasoning process": 27438,
      "parameter model": 24190,
      "thought processes": 33080,
      "complex instructions": 5276,
      "largescale diverse": 17351,
      "surpasses conventional": 31739,
      "conventional stateoftheart": 6079,
      "stateoftheart instructiontuned": 30933,
      "models vicuna13b": 22115,
      "zeroshot reasoning": 35991,
      "reasoning benchmarks": 27380,
      "bigbench hard": 3701,
      "hard bbh": 13421,
      "bbh benchmark": 3297,
      "benchmark shows": 3410,
      "shows competitive": 29924,
      "professional academic": 26175,
      "lsat gre": 19325,
      "zeroshot settings": 35995,
      "gpt4 research": 13112,
      "research indicates": 28327,
      "generated humans": 12364,
      "humans advanced": 13918,
      "promising direction": 26285,
      "model capabilities": 20408,
      "efficient instruction": 9040,
      "instruction optimization": 15172,
      "language modelsllms": 16761,
      "instruction followers": 15160,
      "blackbox llms": 3754,
      "opensource llm": 23517,
      "llm generate": 18313,
      "generate instruction": 12291,
      "using opensource": 34866,
      "zeroshot evaluation": 35968,
      "evaluation performance": 9985,
      "new soft": 22843,
      "improving zeroshot": 14426,
      "different combinations": 8057,
      "llms apis": 18435,
      "apis including": 2107,
      "including vicuna": 14529,
      "chatgpt results": 4494,
      "outperforms sota": 23850,
      "variety downstream": 35060,
      "acquisition children": 1127,
      "children language": 4534,
      "increasingly complex": 14634,
      "remain largely": 27984,
      "largely unknown": 17313,
      "compare learning": 5107,
      "deep language": 7320,
      "training gpt2": 33525,
      "gpt2 exhibits": 12889,
      "scratch evaluate": 29290,
      "step using": 31052,
      "language production": 16805,
      "main findings": 19395,
      "models tend": 22049,
      "tasks learned": 32397,
      "improve training": 14301,
      "models overall": 21788,
      "results shed": 28675,
      "shed new": 29795,
      "new light": 22817,
      "highlight important": 13631,
      "process natural": 26075,
      "reasoning question": 27441,
      "prompts random": 26437,
      "knowledge entities": 15845,
      "entities pretrained": 9635,
      "propose techniques": 26572,
      "specifically use": 30760,
      "prompts guide": 26422,
      "encoded knowledge": 9339,
      "questions random": 27127,
      "paths lead": 24401,
      "applying methods": 2220,
      "questions require": 27130,
      "lossless text": 19256,
      "text compression": 32833,
      "models provide": 21873,
      "provide new": 26718,
      "upper bound": 34482,
      "given window": 12783,
      "past tokens": 24389,
      "significantly smaller": 30086,
      "lossless compression": 19255,
      "preliminary results": 25490,
      "results limited": 28638,
      "experiments suggest": 10486,
      "aims translate": 1673,
      "unified evaluation": 34325,
      "benchmark crosslingual": 3368,
      "comprehensive benchmark": 5356,
      "benchmark study": 3413,
      "models mbart": 21706,
      "models codex": 21060,
      "design experiment": 7703,
      "covering various": 6327,
      "multilingual crosslingual": 22303,
      "zeroshot experiments": 35969,
      "achieve highest": 968,
      "highest performance": 13620,
      "compared popular": 5160,
      "popular models": 25129,
      "performance notably": 24697,
      "models bloom": 21032,
      "training crosslingual": 33463,
      "models mitigated": 21729,
      "fewshot training": 11125,
      "knowledge understanding": 15917,
      "llms achieved": 18412,
      "great success": 13262,
      "general domains": 12162,
      "processing paper": 26124,
      "paper bring": 24018,
      "applications field": 2153,
      "llm research": 18359,
      "llms context": 18490,
      "domain specifically": 8595,
      "specifically train": 30758,
      "corpus including": 6185,
      "supervised data": 31671,
      "domainspecific data": 8648,
      "data construct": 6656,
      "abilities using": 645,
      "using tools": 34929,
      "experiments conducted": 10426,
      "approach datasets": 2255,
      "chatbot arena": 4449,
      "llm based": 18275,
      "based chat": 3142,
      "chat assistants": 4440,
      "existing benchmarks": 10265,
      "human preferences": 13856,
      "preferences address": 25473,
      "explore using": 10604,
      "using strong": 34922,
      "models openended": 21770,
      "llm judges": 18325,
      "platform results": 25009,
      "strong llm": 31181,
      "crowdsourced human": 6433,
      "preferences achieving": 25472,
      "approximate human": 2412,
      "expensive obtain": 10362,
      "variants llama": 35042,
      "llama vicuna": 18151,
      "conversations human": 6108,
      "transfer ability": 33669,
      "gap study": 12111,
      "research questions": 28356,
      "questions does": 27108,
      "models does": 21167,
      "models second": 21958,
      "tasks multilingual": 32422,
      "multilingual reasoning": 22327,
      "types reasoning": 34071,
      "does outperform": 8534,
      "outperform englishcentric": 23772,
      "model furthermore": 20541,
      "types tasks": 34072,
      "exhibit different": 10213,
      "transfer abilities": 33668,
      "findings demonstrate": 11233,
      "models possess": 21824,
      "experiments provide": 10470,
      "insights enhancing": 15073,
      "2023 shared": 234,
      "paper describes": 24032,
      "task generation": 32133,
      "task benchmark": 32085,
      "models act": 20956,
      "including alpaca": 14455,
      "flant5 gpt2": 11593,
      "evaluated terms": 9883,
      "ability based": 657,
      "automated human": 2863,
      "evaluation generated": 9955,
      "generated responses": 12385,
      "gpt35 using": 13038,
      "using ensemble": 34772,
      "given dialogue": 12746,
      "dialogue contexts": 8013,
      "participating teams": 24332,
      "results highlight": 28622,
      "highlight need": 13634,
      "better suited": 3628,
      "translation large": 33830,
      "models nonenglish": 21757,
      "content analysis": 5851,
      "analysis recent": 1955,
      "years large": 35891,
      "models open": 21767,
      "gpt4 metas": 13095,
      "llama googles": 18108,
      "googles palm": 12833,
      "dominant approach": 8660,
      "approach building": 2244,
      "generate language": 12298,
      "content moderation": 5864,
      "languages recently": 16912,
      "recently researchers": 27621,
      "technology companies": 32680,
      "extend capabilities": 10647,
      "explore capabilities": 10576,
      "explanation large": 10537,
      "data english": 6692,
      "languages multilingual": 16900,
      "models attempt": 20992,
      "analysis large": 1934,
      "companies researchers": 5069,
      "developing deploying": 7941,
      "design implementation": 7705,
      "generative artificial": 12653,
      "emergence large": 9169,
      "ai agents": 1595,
      "paper aim": 24008,
      "llms telecom": 18998,
      "telecom domain": 32683,
      "finetune llms": 11294,
      "including bert": 14461,
      "languages demonstrate": 16869,
      "demonstrate use": 7508,
      "use case": 34515,
      "selected models": 29382,
      "demonstrate finetuning": 7457,
      "finetuning bert": 11377,
      "roberta model": 28921,
      "accuracy gpt2": 886,
      "model 50": 20336,
      "similar performance": 30113,
      "pretrained llm": 25705,
      "llm effectively": 18293,
      "developed framework": 7927,
      "paves way": 24419,
      "capabilities natural": 4047,
      "pose significant": 25161,
      "significant risks": 30022,
      "harmful text": 13444,
      "suite opensource": 31600,
      "opensource code": 23489,
      "code repositories": 4810,
      "llms based": 18443,
      "transformers gpts": 33784,
      "goal project": 12808,
      "create worlds": 6357,
      "worlds best": 35837,
      "opensource alternative": 23484,
      "opensource community": 23494,
      "opensource finetuned": 23502,
      "commercial use": 4998,
      "apache 20": 2098,
      "release 100": 27899,
      "opensource language": 23506,
      "ai development": 1607,
      "models needs": 21749,
      "llms structured": 18971,
      "recent months": 27536,
      "potential artificial": 25240,
      "weights public": 35512,
      "demonstrating impressive": 7582,
      "capabilities generative": 4022,
      "lms believe": 19071,
      "lms solving": 19110,
      "solving tasks": 30518,
      "analysis providing": 1952,
      "problemsolving paper": 26040,
      "received little": 27478,
      "little attention": 18049,
      "attention present": 2736,
      "new algorithm": 22773,
      "lms use": 19120,
      "program execution": 26194,
      "model hope": 20565,
      "light need": 17829,
      "crosslingual alignment": 6413,
      "alignment instruction": 1763,
      "foundation llms": 11795,
      "llms instructionfollowing": 18725,
      "instructionfollowing llms": 15234,
      "llms instruction": 18723,
      "plays vital": 25032,
      "vital role": 35371,
      "aligning llms": 1749,
      "preferences existing": 25474,
      "llms usually": 19030,
      "focused english": 11662,
      "performance nonenglish": 24694,
      "nonenglish languages": 22992,
      "languages order": 16902,
      "llms construct": 18489,
      "human workload": 13878,
      "propose transfer": 26577,
      "capabilities language": 4029,
      "translation task": 33854,
      "llama foundation": 18103,
      "llm automatically": 18272,
      "automatically constructing": 2908,
      "translation instructions": 33828,
      "despite utilizing": 7826,
      "considerably smaller": 5713,
      "smaller parameter": 30394,
      "results translation": 28699,
      "translation tasks": 33855,
      "translation capability": 33822,
      "compared gpt4": 5138,
      "gpt4 automatic": 13055,
      "estimate performance": 9781,
      "performance general": 24610,
      "general tasks": 12187,
      "instruction test": 15180,
      "achieves 89": 1025,
      "performance knowledge": 24638,
      "available finetuning": 2975,
      "scientific writing": 29261,
      "writing support": 35859,
      "trained corpus": 33387,
      "corpus scientific": 6187,
      "score indicates": 29271,
      "investigate effect": 15580,
      "potential biases": 25245,
      "sentence likely": 29537,
      "finally propose": 11201,
      "alternative given": 1853,
      "word substitutions": 35650,
      "writing style": 35857,
      "using context": 34758,
      "produce output": 26152,
      "gold standard": 12814,
      "t5 large": 31953,
      "considering various": 5726,
      "input sentence": 15025,
      "aims explain": 1664,
      "capabilities deep": 4009,
      "gradientbased training": 13194,
      "gap theory": 12112,
      "theory practice": 33060,
      "trajectory arbitrary": 33661,
      "broad range": 3885,
      "networks transformers": 22716,
      "training algorithms": 33440,
      "sgd adam": 29771,
      "exploit lowrank": 10555,
      "design new": 7711,
      "new training": 22861,
      "training propose": 33594,
      "total training": 33300,
      "counterfactual data": 6295,
      "finegrained sentiment": 11277,
      "evaluation aspects": 9921,
      "great impact": 13249,
      "impact models": 14133,
      "performance mitigate": 24678,
      "mitigate problem": 20256,
      "novel simple": 23108,
      "method generate": 19924,
      "opinion expressions": 23578,
      "sentiment polarity": 29571,
      "integrated gradients": 15325,
      "original text": 23721,
      "text pretrained": 32917,
      "model plm": 20712,
      "plm t5": 25037,
      "shows proposed": 29935,
      "method performs": 19955,
      "better current": 3598,
      "current augmentation": 6483,
      "new large": 22813,
      "smaller size": 30397,
      "transformerbased model": 33759,
      "model 13b": 20335,
      "quality data": 26949,
      "1b tokens": 199,
      "despite small": 7815,
      "pass1 accuracy": 24373,
      "finetuning stage": 11534,
      "350m parameters": 347,
      "trained pipeline": 33417,
      "deep fusion": 7318,
      "remarkable progress": 28054,
      "progress wide": 26229,
      "range domains": 27191,
      "domains particularly": 8634,
      "llms need": 18815,
      "need large": 22634,
      "resources time": 28447,
      "offer potential": 23292,
      "potential cost": 25250,
      "underlying mechanisms": 34159,
      "poorly understood": 25108,
      "present notable": 25543,
      "efficient approach": 9025,
      "leverages pretrained": 17772,
      "second propose": 29326,
      "theoretical framework": 33049,
      "framework using": 11903,
      "error analysis": 9708,
      "practical effective": 25364,
      "effective approach": 8860,
      "process reduces": 26079,
      "reduces computational": 27733,
      "computational requirements": 5476,
      "traditional training": 33353,
      "finally validate": 11206,
      "framework guides": 11863,
      "optimal use": 23617,
      "use deep": 34524,
      "training dynamics": 33505,
      "reduces training": 27740,
      "llms express": 18595,
      "empowering large": 9269,
      "methods primarily": 20077,
      "primarily rely": 25912,
      "access internal": 817,
      "internal model": 15439,
      "model information": 20582,
      "information model": 14889,
      "llms especially": 18567,
      "growing need": 13315,
      "need explore": 22627,
      "approaches llm": 2381,
      "uncertainty estimation": 34118,
      "systematic framework": 31873,
      "sampling methods": 29094,
      "methods generating": 20041,
      "multiple responses": 22415,
      "methods key": 20055,
      "types datasets": 34059,
      "widelyused llms": 35579,
      "including gpt4": 14490,
      "gpt4 llama": 13089,
      "llama chat": 18082,
      "analysis uncovers": 1975,
      "key insights": 15775,
      "tend overconfident": 32708,
      "model capability": 20409,
      "performance improve": 24624,
      "proposed strategies": 26620,
      "responses better": 28486,
      "help mitigate": 13509,
      "techniques consistently": 32632,
      "consistently outperform": 5754,
      "challenging tasks": 4403,
      "professional knowledge": 26176,
      "indicating significant": 14704,
      "study serve": 31396,
      "serve strong": 29649,
      "provide insights": 26709,
      "arithmetic operations": 2503,
      "efficient alternative": 9024,
      "finetuning parameterefficient": 11470,
      "dataset underlying": 7050,
      "underlying pretrained": 34161,
      "model remains": 20752,
      "remains unchanged": 28015,
      "representing diverse": 28195,
      "applied various": 2200,
      "various domains": 35087,
      "weight space": 35497,
      "capabilities specifically": 4072,
      "addition negation": 1243,
      "approach requires": 2335,
      "highly flexible": 13664,
      "apply different": 2206,
      "domain transfer": 8600,
      "instructiontuned large": 15286,
      "based llama": 3191,
      "llama empirical": 18094,
      "demonstrate approach": 7436,
      "approach produces": 2327,
      "existing ones": 10302,
      "harnessing llms": 13466,
      "design using": 7717,
      "using gpt4": 34792,
      "learning objectives": 17624,
      "transformer gpt4": 33722,
      "generation explanation": 12499,
      "emerging technology": 9198,
      "study models": 31366,
      "course design": 6311,
      "focus specific": 11657,
      "specific cognitive": 30685,
      "time consuming": 33114,
      "carefully crafted": 4174,
      "gpt4 conceptual": 13068,
      "best practices": 3575,
      "analysis showed": 1964,
      "models support": 22032,
      "texttoimage generation": 33006,
      "generation lightweight": 12540,
      "pure language": 26880,
      "pythia gpt2": 26900,
      "qa reasoning": 26917,
      "generation machine": 12543,
      "translation achieving": 33818,
      "model twice": 20845,
      "methods like": 20060,
      "like chainofthought": 17852,
      "tasks used": 32542,
      "increase faithfulness": 14596,
      "prompts human": 26423,
      "tensortrain decomposition": 32727,
      "llms capture": 18459,
      "significantly enhance": 30047,
      "complex language": 5277,
      "language patterns": 16771,
      "parameters prohibitively": 24280,
      "model storage": 20809,
      "issue work": 15663,
      "proposes approach": 26625,
      "approach based": 2242,
      "token embedding": 33189,
      "distributed manner": 8385,
      "manner experimental": 19547,
      "gpt2 demonstrate": 12881,
      "performance original": 24705,
      "remarkable potential": 28052,
      "potential natural": 25281,
      "challenge lies": 4317,
      "susceptibility hallucinations": 31781,
      "uncertainty quantification": 34119,
      "presents promising": 25591,
      "llms remains": 18905,
      "significant hurdle": 29986,
      "address critical": 1319,
      "llmgenerated text": 18395,
      "tokens carry": 33219,
      "phenomenon linguistic": 24924,
      "existing methodologies": 10292,
      "methodologies treat": 19990,
      "estimating uncertainty": 9785,
      "propose jointly": 26521,
      "attention relevant": 2739,
      "experiments involving": 10453,
      "range popular": 27208,
      "popular offtheshelf": 25131,
      "offtheshelf llms": 23329,
      "instructiontuned llms": 15296,
      "llms vicuna": 19037,
      "vicuna wizardlm": 35257,
      "like opt": 17891,
      "opt llama": 23596,
      "33b parameters": 333,
      "evaluation various": 10025,
      "tasks encompassing": 32310,
      "science qa": 29243,
      "medical qa": 19770,
      "addressing challenges": 1370,
      "memory efficient": 19814,
      "demonstrated excellent": 7519,
      "excellent performance": 10162,
      "high cost": 13561,
      "memory overheads": 19824,
      "memory usage": 19834,
      "performance penalty": 24714,
      "efficient optimizers": 9052,
      "fast convergence": 10991,
      "methods extensive": 20031,
      "demonstrate training": 7507,
      "training stability": 33621,
      "tasks bert": 32252,
      "training notably": 33576,
      "faster convergence": 10997,
      "available multilingual": 2990,
      "source knowledge": 30561,
      "information easily": 14863,
      "easily accessible": 8791,
      "texts contain": 32990,
      "contain complex": 5826,
      "convey information": 6117,
      "extracted pretrained": 10751,
      "results wellknown": 28710,
      "evaluation approach": 9920,
      "competitively compared": 5233,
      "participating systems": 24331,
      "metrics model": 20144,
      "alignment efficient": 1758,
      "unified model": 34332,
      "llms typically": 19014,
      "nextword prediction": 22899,
      "formulation tasks": 11777,
      "tasks demanding": 32285,
      "suboptimal performance": 31439,
      "build efficient": 3909,
      "sizes paper": 30302,
      "model wide": 20870,
      "factual consistency": 10878,
      "texts model": 32997,
      "alignment model": 1773,
      "finetuning roberta": 11515,
      "355m parameters": 349,
      "datasets despite": 7096,
      "experiments models": 10462,
      "20 datasets": 209,
      "tasks model": 32419,
      "model matches": 20636,
      "matches surpasses": 19657,
      "flant5 models": 11600,
      "single unified": 30228,
      "outperforms taskspecific": 23860,
      "applied evaluate": 2188,
      "evaluate factual": 9837,
      "consistency language": 5732,
      "model improves": 20574,
      "improves various": 14401,
      "various baselines": 35076,
      "gpt35 chatgpt": 13018,
      "lightweight model": 17842,
      "model serve": 20774,
      "llms gpt35": 18649,
      "answering tasks": 2077,
      "improving average": 14404,
      "exact match": 10088,
      "match em": 19640,
      "em score": 9119,
      "questions evaluating": 27109,
      "evaluating effectiveness": 9893,
      "textual descriptions": 33027,
      "utilize llms": 34960,
      "embeddings preserve": 9144,
      "challenges remain": 4374,
      "research highlights": 28322,
      "highlights need": 13653,
      "need improvement": 22632,
      "improvement terms": 14348,
      "terms capturing": 32740,
      "support various": 31713,
      "models generating": 21313,
      "generating accurate": 12408,
      "use nlp": 34556,
      "educational applications": 8839,
      "task study": 32195,
      "study attempt": 31302,
      "abilities large": 628,
      "present extensive": 25531,
      "fewshot incontext": 11107,
      "finetuned flant5": 11309,
      "using reinforcement": 34900,
      "learning experimental": 17565,
      "experimental findings": 10387,
      "indicate efficacy": 14687,
      "models measured": 21711,
      "significant challenges": 29966,
      "challenges finetuning": 4345,
      "models finally": 21270,
      "skills large": 30310,
      "perspectives different": 24900,
      "llms exhibit": 18578,
      "humans tend": 13932,
      "contexts introduce": 5941,
      "introduce concept": 15501,
      "refers models": 27778,
      "experiments use": 10495,
      "based different": 3152,
      "different perspectives": 8120,
      "experiments llms": 10461,
      "implicitly explicitly": 14179,
      "prompt llms": 26334,
      "conduct quantitative": 5616,
      "quantitative experiments": 26993,
      "different models": 8109,
      "models gpt4": 21332,
      "gpt4 gpt35": 13082,
      "effectiveness various": 8971,
      "scientific questions": 29255,
      "project website": 26247,
      "website available": 35488,
      "models current": 21118,
      "developments large": 7980,
      "llms enabled": 18561,
      "impressive zeroshot": 14250,
      "zeroshot capabilities": 35954,
      "highly challenging": 13659,
      "abilities llms": 633,
      "relative comparisons": 27877,
      "work examines": 35702,
      "multiple perspectives": 22408,
      "general effective": 12164,
      "llms flant5": 18615,
      "flant5 llama2chat": 11598,
      "achieve performance": 977,
      "performance competitive": 24554,
      "competitive stateoftheart": 5230,
      "stateoftheart methods": 30954,
      "methods additionally": 20004,
      "additionally demonstrate": 1278,
      "exhibit strong": 10229,
      "debiasing methods": 7216,
      "methods improve": 20046,
      "policy improve": 25084,
      "models math": 21705,
      "math problems": 19670,
      "solve math": 30492,
      "problems language": 26028,
      "sampling strategy": 29097,
      "math reasoning": 19672,
      "reasoning step": 27453,
      "generate wrong": 12339,
      "abstract level": 769,
      "token probability": 33199,
      "select token": 29380,
      "test method": 32774,
      "gsm8k dataset": 13326,
      "dataset gpt2": 6998,
      "performance gain": 24605,
      "implementation available": 14161,
      "planning long": 25002,
      "long context": 19167,
      "context understanding": 5924,
      "llms recently": 18894,
      "recently achieved": 27586,
      "achieved better": 1003,
      "web automation": 35474,
      "performance realworld": 24732,
      "limited context": 17946,
      "inductive bias": 14738,
      "agent learns": 1562,
      "tasks real": 32471,
      "following natural": 11698,
      "python programs": 26904,
      "llms long": 18782,
      "documents using": 8520,
      "global attention": 12795,
      "attention mechanisms": 2727,
      "denoising objectives": 7601,
      "solve various": 30498,
      "tasks achieving": 32232,
      "higher success": 13606,
      "models graphtotext": 21337,
      "graphtotext generation": 13241,
      "llms widely": 19046,
      "widely employed": 35573,
      "process finetuning": 26060,
      "llms requires": 18911,
      "significant training": 30028,
      "training resources": 33599,
      "explore capability": 10577,
      "generate descriptive": 12270,
      "graph data": 13220,
      "data zeroshot": 6917,
      "specifically evaluate": 30737,
      "llm models": 18339,
      "generating fluent": 12423,
      "fluent coherent": 11633,
      "achieving bleu": 1087,
      "bleu scores": 3769,
      "struggle understanding": 31247,
      "semantic relations": 29466,
      "tend generate": 32705,
      "irrelevant information": 15644,
      "detect machinegenerated": 7850,
      "scores text": 29283,
      "generated generative": 12357,
      "transportation safety": 33866,
      "remarkable effectiveness": 28041,
      "various generaldomain": 35096,
      "generaldomain natural": 12191,
      "domain tasks": 8598,
      "tasks suboptimal": 32517,
      "primarily attributed": 25908,
      "accurate responses": 928,
      "responses address": 28485,
      "address challenge": 1308,
      "challenge introduce": 4316,
      "instructionoutput pairs": 15238,
      "dataset accessible": 6935,
      "key technology": 15788,
      "development large": 7957,
      "llms involves": 18730,
      "align models": 1735,
      "models responses": 21939,
      "responses human": 28496,
      "human expectations": 13813,
      "major approaches": 19436,
      "finetuning sft": 11520,
      "sft reinforcement": 29764,
      "feedback rlhf": 11071,
      "produce best": 26139,
      "commercial llms": 4988,
      "improve accessibility": 14255,
      "llms research": 18915,
      "research development": 28307,
      "llms introduced": 18728,
      "alpaca vicuna": 1833,
      "existing opensource": 10304,
      "llms instructiontuned": 18726,
      "popular languages": 25118,
      "tuning llms": 33996,
      "significant gap": 29983,
      "diverse languages": 8436,
      "important questions": 14209,
      "overcome issue": 23921,
      "introduces instruction": 15542,
      "development future": 7954,
      "multilingual llm": 22317,
      "research present": 28345,
      "present benchmark": 25515,
      "evaluation generative": 9957,
      "languages experiments": 16874,
      "demonstrate advantages": 7434,
      "sft different": 29759,
      "different base": 8054,
      "sentence embeddings": 29534,
      "learning llms": 17599,
      "tasks application": 32244,
      "area ongoing": 2482,
      "ongoing research": 23358,
      "propose incontext": 26518,
      "incontext learningbased": 14567,
      "learningbased method": 17675,
      "aimed improving": 1654,
      "approach involves": 2302,
      "previous promptbased": 25873,
      "representation method": 28145,
      "enables llms": 9303,
      "learning scaling": 17649,
      "experiments incontext": 10449,
      "learning enables": 17563,
      "finetuning helps": 11415,
      "performance comparable": 24547,
      "methods scaling": 20091,
      "size scaling": 30283,
      "largest model": 17396,
      "stateoftheart result": 30982,
      "tasks finetune": 32334,
      "llms current": 18501,
      "opt model": 23599,
      "promptbased method": 26362,
      "method surpasses": 19978,
      "surpasses performance": 31748,
      "llms moral": 18804,
      "reasoning large": 27415,
      "tasks wide": 32549,
      "range different": 27190,
      "important know": 14203,
      "issues paper": 15669,
      "employ methods": 9254,
      "gemini pro": 12141,
      "anthropics claude": 2090,
      "openais gpt4": 23446,
      "alignment human": 1762,
      "human responses": 13861,
      "alignment gpt4": 1761,
      "lead models": 17466,
      "llmgenerated responses": 18394,
      "responses highly": 28495,
      "correlated human": 6214,
      "human participants": 13849,
      "need research": 22641,
      "opensource large": 23509,
      "specially designed": 30681,
      "generation llms": 12541,
      "training llms": 33554,
      "dataset collected": 6951,
      "title abstract": 33178,
      "developed finetuning": 7926,
      "general llms": 12177,
      "demonstrated powerful": 7540,
      "field experiments": 11136,
      "academic papers": 787,
      "comparable chatgpt": 5077,
      "outperforms opensource": 23840,
      "llama13b model": 18157,
      "led paradigm": 17686,
      "primary objective": 25923,
      "assess effectiveness": 2591,
      "effectiveness models": 8961,
      "prompting models": 26390,
      "tasks past": 32448,
      "aim evaluate": 1641,
      "performance current": 24560,
      "current large": 6502,
      "models highlighting": 21357,
      "constraints context": 5790,
      "largescale synthetic": 17381,
      "dataset related": 7031,
      "vqa dataset": 35396,
      "build dataset": 3908,
      "context provided": 5910,
      "palm2 paper": 23996,
      "models llava": 21468,
      "llava mplugowl": 18249,
      "cider score": 4566,
      "accuracy improvement": 890,
      "mask tokens": 19608,
      "leveraging larger": 17787,
      "prompting techniques": 26398,
      "speculative decoding": 30778,
      "advances large": 1478,
      "capabilities propose": 4065,
      "accelerate llm": 791,
      "scenarios address": 29201,
      "improving previous": 14419,
      "second stage": 29327,
      "smaller language": 30376,
      "rationale generation": 27277,
      "strong reasoning": 31190,
      "evaluate methods": 9849,
      "methods improvement": 20047,
      "rationales generated": 27279,
      "generated larger": 12371,
      "model longer": 20631,
      "longer contexts": 19197,
      "retrieval method": 28745,
      "involves training": 15632,
      "score generated": 29268,
      "generated rationales": 12383,
      "retrieved contexts": 28772,
      "knowledge sources": 15909,
      "second method": 29325,
      "relevant information": 27942,
      "improve results": 14294,
      "improves strong": 14395,
      "type question": 34055,
      "question selecting": 27074,
      "does better": 8524,
      "proposed models": 26614,
      "generally outperform": 12243,
      "stablevicuna 13b": 30853,
      "fewshot chainofthought": 11099,
      "existing large": 10283,
      "imbalance training": 14101,
      "english tasks": 9492,
      "tasks languages": 32389,
      "semantic alignment": 29447,
      "general task": 12185,
      "crosslingual models": 6417,
      "scaling laws": 29172,
      "investigate advantages": 15575,
      "perform multilingual": 24494,
      "resources build": 28430,
      "resourceconstrained setting": 28422,
      "alpaca average": 1824,
      "translation dataset": 33824,
      "outperform previous": 23782,
      "llamabased models": 18238,
      "models average": 21004,
      "demonstrates ability": 7556,
      "response content": 28476,
      "representation space": 28150,
      "middle layers": 20159,
      "tasks ecommerce": 32304,
      "recently instructionfollowing": 27603,
      "instructionfollowing large": 15230,
      "represented chatgpt": 28191,
      "exhibited exceptional": 10234,
      "unique characteristics": 34357,
      "ecommerce data": 8809,
      "llm tailored": 18372,
      "solve issue": 30491,
      "instruction dataset": 15150,
      "information user": 14923,
      "user reviews": 34670,
      "final task": 11185,
      "different parameter": 8117,
      "parameter scales": 24196,
      "backbone model": 3055,
      "model bloomz": 20403,
      "exhibits excellent": 10246,
      "capabilities extensive": 4017,
      "evaluations demonstrate": 10027,
      "outperforms chatgpt": 23812,
      "improved loss": 14314,
      "writing natural": 35853,
      "neural code": 22723,
      "techniques generating": 32640,
      "generating descriptions": 12418,
      "descriptions using": 7692,
      "propose evaluate": 26509,
      "similarity metric": 30130,
      "loss output": 19250,
      "output sentence": 23880,
      "training batch": 33443,
      "compared baselines": 5123,
      "report improvement": 28116,
      "vast majority": 35187,
      "trajectories using": 33659,
      "tokens using": 33251,
      "using architecture": 34733,
      "gpt2 sequence": 12949,
      "model sequentially": 20772,
      "special tokens": 30657,
      "tokens training": 33250,
      "training tokens": 33636,
      "increasing use": 14631,
      "use internet": 34538,
      "created comprehensive": 6359,
      "detection model": 7872,
      "editing model": 8829,
      "approach utilizes": 2358,
      "model controlled": 20444,
      "achieves average": 1029,
      "average bleu": 3011,
      "gpt2 roberta": 12948,
      "dataset achieve": 6936,
      "field previous": 11150,
      "previous attempts": 25866,
      "models optimization": 21779,
      "behavior large": 3315,
      "models pressing": 21838,
      "finetuning reinforcement": 11503,
      "forward pass": 11784,
      "steering vectors": 31030,
      "gpt2 openwebtext": 12930,
      "approach yields": 2364,
      "properties output": 26478,
      "method requires": 19968,
      "language specification": 16823,
      "enhancing reliability": 9574,
      "models emergence": 21188,
      "showcasing exceptional": 29849,
      "tasks existing": 32320,
      "existing research": 10313,
      "transformers like": 33791,
      "applicability findings": 2119,
      "paper embarks": 24037,
      "empirical investigation": 9228,
      "domain llms": 8574,
      "llms focusing": 18616,
      "focusing llama": 11672,
      "thoroughly evaluate": 33076,
      "finetuning scenarios": 11517,
      "scenarios notably": 29212,
      "finetuning generative": 11410,
      "objective llms": 23207,
      "tasks findings": 32332,
      "cosine distance": 6240,
      "demonstrates superior": 7576,
      "superior efficacy": 31647,
      "provide intriguing": 26712,
      "explanation phenomenon": 10538,
      "embedding spaces": 9133,
      "bert family": 3502,
      "enhances understanding": 9553,
      "llms detect": 18534,
      "data enhancing": 6694,
      "enhancing adaptability": 9556,
      "dynamic environments": 8757,
      "evaluation nlp": 9982,
      "educational materials": 8842,
      "fields like": 11156,
      "expensive create": 10359,
      "achieved significant": 1016,
      "success various": 31528,
      "education domain": 8835,
      "explored work": 10611,
      "work examine": 35701,
      "nlp computer": 22928,
      "automated benchmarks": 2856,
      "benchmarks reveal": 3472,
      "reveal gpt4": 28798,
      "like gpt35": 17871,
      "gpt35 palm2": 13033,
      "palm2 llama2": 23995,
      "compare human": 5106,
      "gptbased evaluation": 13140,
      "evaluation scores": 10006,
      "provide indepth": 26707,
      "analysis findings": 1921,
      "ones certain": 23345,
      "certain limitations": 4276,
      "limitations observed": 17928,
      "content occasionally": 5865,
      "factual errors": 10882,
      "errors compared": 9722,
      "systematic bias": 31868,
      "using gpt": 34787,
      "gpt evaluation": 12846,
      "closing gap": 4701,
      "foreign languages": 11725,
      "model scratch": 20768,
      "language domain": 16067,
      "serves essential": 29654,
      "particular linguistic": 24340,
      "domain context": 8558,
      "context ii": 5893,
      "approach lies": 2311,
      "size number": 30267,
      "llama llama2": 18120,
      "scenarios involving": 29210,
      "memory resources": 19827,
      "tokens required": 33243,
      "required represent": 28232,
      "present methodology": 25540,
      "methodology named": 19997,
      "research demonstrates": 28302,
      "methodology applied": 19992,
      "architecture model": 2447,
      "model known": 20599,
      "resulting model": 28557,
      "results significant": 28683,
      "significant reduction": 30019,
      "reduction number": 27764,
      "tasks achieved": 32231,
      "model compared": 20431,
      "compared traditional": 5177,
      "pretraining approach": 25784,
      "7b models": 537,
      "models english": 21204,
      "english pretrained": 9489,
      "foundational language": 11803,
      "models foundational": 21296,
      "advanced natural": 1433,
      "researchers developing": 28375,
      "contextualized language": 5961,
      "language representation": 16817,
      "potential smaller": 25299,
      "benchmark realworld": 3407,
      "using datasets": 34766,
      "mental health": 19839,
      "present strong": 25554,
      "strong general": 31171,
      "capabilities current": 4007,
      "specialized capabilities": 30667,
      "tuning standard": 34015,
      "instruction input": 15170,
      "mechanism llms": 19751,
      "llms limitations": 18765,
      "focus llms": 11651,
      "llms tend": 19000,
      "tend focus": 32704,
      "instructionfollowing dataset": 15227,
      "model instruction": 20586,
      "instruction understanding": 15210,
      "translation apply": 33819,
      "apply methods": 2210,
      "methods mainstream": 20066,
      "bloom llama": 3786,
      "demonstrate significant": 7492,
      "improvements translation": 14366,
      "average improvement": 3019,
      "english german": 9481,
      "metric based": 20119,
      "based word": 3238,
      "recently growing": 27600,
      "benchmarks proposed": 3469,
      "evaluate ability": 9817,
      "ability llms": 697,
      "data leakage": 6754,
      "ability paper": 708,
      "issues based": 15667,
      "based blooms": 3140,
      "particular design": 24336,
      "potential data": 25251,
      "leakage objective": 17499,
      "objective subjective": 23208,
      "evaluation llms": 9970,
      "llms comprehensive": 18480,
      "comprehensive experiments": 5376,
      "experiments advanced": 10417,
      "advanced llms": 1430,
      "gpt4 achieves": 13053,
      "achieves sota": 1067,
      "llms substantial": 18975,
      "substantial room": 31476,
      "room improvement": 28967,
      "improvement especially": 14337,
      "data codes": 6638,
      "codes publicly": 4854,
      "instruction tuned": 15183,
      "models advent": 20969,
      "advent large": 1509,
      "processing enabling": 26101,
      "progress various": 26228,
      "various applications": 35074,
      "structured information": 31221,
      "models focus": 21288,
      "llama architecture": 18076,
      "wikipedia dataset": 35603,
      "low rank": 19272,
      "rank adaptation": 27222,
      "lora technique": 19235,
      "dense passage": 7608,
      "passage retrieval": 24380,
      "answer relevant": 2052,
      "entity relation": 9652,
      "model achieved": 20345,
      "achieved average": 1000,
      "average f1": 3015,
      "generation recommendation": 12593,
      "recommendation paper": 27649,
      "suggestions based": 31587,
      "text prompt": 32922,
      "model extract": 20509,
      "new features": 22803,
      "features users": 11044,
      "user study": 34675,
      "study comparing": 31308,
      "generated finetuned": 12356,
      "model outperformed": 20669,
      "fedllm using": 11053,
      "forward gradient": 11783,
      "federated learning": 11050,
      "learning fl": 17570,
      "user data": 34645,
      "data privacy": 6807,
      "tasks approach": 32246,
      "vast model": 35188,
      "challenges concerning": 4339,
      "mobile devices": 20319,
      "significant memory": 29996,
      "memory consumption": 19809,
      "model convergence": 20445,
      "response challenges": 28475,
      "work introduces": 35726,
      "designed enhance": 7729,
      "key idea": 15770,
      "methods requiring": 20090,
      "memory efficiency": 19813,
      "efficiency time": 9016,
      "time efficiency": 33119,
      "key designs": 15762,
      "parameterefficient training": 24214,
      "methods essential": 20027,
      "approach llm": 2312,
      "speed accuracy": 30795,
      "valuable model": 35015,
      "llms nlp": 18818,
      "significant advantages": 29959,
      "conventional methods": 6075,
      "orders magnitude": 23681,
      "magnitude faster": 19382,
      "reduction memory": 27763,
      "memory footprint": 19815,
      "loss functions": 19245,
      "advancements natural": 1470,
      "models computationally": 21083,
      "techniques reduce": 32658,
      "reduce size": 27725,
      "student model": 31254,
      "explore various": 10605,
      "various techniques": 35152,
      "transformer layer": 33727,
      "methods tuning": 20105,
      "proposed techniques": 26622,
      "goal work": 12809,
      "improve efficiency": 14264,
      "efficiency effectiveness": 9002,
      "accurate models": 926,
      "optimization using": 23637,
      "important challenge": 14198,
      "specifically focus": 30742,
      "domain specific": 8594,
      "end propose": 9416,
      "deep reinforcement": 7343,
      "learning rl": 17646,
      "rl based": 28901,
      "based search": 3220,
      "search optimal": 29310,
      "deep rl": 7345,
      "performance open": 24700,
      "observe average": 23225,
      "diverse benchmark": 8414,
      "benchmark including": 3392,
      "proposed approach": 26590,
      "research recent": 28357,
      "develop endtoend": 7915,
      "depends heavily": 7624,
      "diversity quality": 8482,
      "emergence powerful": 9179,
      "promising avenue": 26284,
      "proposes novel": 26630,
      "domain generates": 8567,
      "instructions based": 15244,
      "based input": 3178,
      "transformer decoder": 33713,
      "showcase practical": 29839,
      "applications benefit": 2142,
      "integrating large": 15331,
      "enable automated": 9284,
      "findings validate": 11261,
      "validate efficacy": 34995,
      "efficacy proposed": 8991,
      "future advancements": 12028,
      "intelligence large": 15358,
      "language learning": 16109,
      "progress large": 26212,
      "llms impressive": 18685,
      "efficacy realworld": 8993,
      "expert knowledge": 10511,
      "remains unclear": 28016,
      "hold potential": 13736,
      "development artificial": 7949,
      "ai based": 1600,
      "evaluating efficacy": 9894,
      "second language": 29322,
      "multiplechoice question": 22430,
      "effectiveness llms": 8957,
      "including understanding": 14526,
      "understanding application": 34208,
      "language knowledge": 16104,
      "influence various": 14842,
      "various prompting": 35135,
      "techniques zero": 32674,
      "fewshot method": 11116,
      "cot think": 6284,
      "think stepbystep": 33066,
      "external tools": 10736,
      "conducted largescale": 5639,
      "popular llms": 25123,
      "distinct models": 8367,
      "using methods": 34840,
      "improvements compared": 14357,
      "compared zeroshot": 5184,
      "zeroshot baseline": 35953,
      "practical questions": 25370,
      "reasoning realworld": 27446,
      "additionally explore": 1284,
      "preliminary findings": 25489,
      "user interaction": 34657,
      "approaches suffer": 2398,
      "suffer poor": 31553,
      "limited language": 17955,
      "manual efforts": 19558,
      "understanding reasoning": 34265,
      "perspective task": 24898,
      "unified language": 34330,
      "capable handling": 4108,
      "arbitrary tasks": 2431,
      "key insight": 15774,
      "llms domainspecific": 18547,
      "domainspecific knowledge": 8651,
      "main components": 19392,
      "memory injection": 19820,
      "injection techniques": 14983,
      "performance new": 24691,
      "common tasks": 5017,
      "results demonstrated": 28598,
      "openais gpt": 23440,
      "marked significant": 19593,
      "significant advancement": 29950,
      "trained vast": 33435,
      "amounts text": 1888,
      "data llms": 6759,
      "llms capable": 18458,
      "capable understanding": 4122,
      "generating humanlike": 12428,
      "range topics": 27214,
      "expands applications": 10343,
      "applications llms": 2167,
      "llms exploring": 18593,
      "data preprocessing": 6803,
      "data mining": 6769,
      "analytics applications": 1986,
      "detection data": 7864,
      "tasks alongside": 32240,
      "inherent capabilities": 14948,
      "llms highlight": 18672,
      "limitations particularly": 17929,
      "particularly terms": 24356,
      "terms computational": 32741,
      "propose llmbased": 26525,
      "llmbased framework": 18388,
      "engineering techniques": 9471,
      "models effectiveness": 21183,
      "experimental study": 10410,
      "12 datasets": 89,
      "datasets gpt4": 7127,
      "achieving 100": 1080,
      "100 accuracy": 33,
      "accuracy f1": 879,
      "score datasets": 29267,
      "immense potential": 14115,
      "study underscores": 31403,
      "underscores promise": 34183,
      "llms domain": 18546,
      "future developments": 12030,
      "specific models": 30707,
      "instead individual": 15122,
      "usually suffer": 34948,
      "model search": 20769,
      "limited number": 17958,
      "address problems": 1351,
      "harness inherent": 13453,
      "update scheme": 34474,
      "enables training": 9310,
      "dynamic model": 8761,
      "storage requirement": 31077,
      "validated various": 35001,
      "llama bert": 18080,
      "image classification": 14061,
      "classification demonstrating": 4595,
      "demonstrating superiority": 7592,
      "superiority existing": 31663,
      "able train": 756,
      "awareness llms": 3045,
      "llms model": 18803,
      "safety alignment": 29042,
      "reasoning contrast": 27397,
      "finetune llm": 11293,
      "examples demonstrations": 10120,
      "assess model": 2598,
      "llms succeed": 18977,
      "size findings": 30248,
      "llms code": 18468,
      "framework pretraining": 11888,
      "models limited": 21462,
      "limited resources": 17962,
      "computational demands": 5465,
      "demands hinder": 7415,
      "large portion": 17258,
      "challenge present": 4326,
      "framework efficient": 11841,
      "models drawing": 21175,
      "t5base model": 31970,
      "gpu just": 13172,
      "loss performance": 19251,
      "modelling research": 20914,
      "implementations make": 14166,
      "technical report": 32607,
      "report large": 28117,
      "way interact": 35436,
      "information conduct": 14856,
      "llms remain": 18904,
      "progress opensource": 26223,
      "limited ability": 17937,
      "longer sequence": 19200,
      "sequence lengths": 29602,
      "context address": 5880,
      "series 7b": 29632,
      "7b parameter": 540,
      "models 8k": 20930,
      "tokens finetuned": 33227,
      "instructional data": 15212,
      "commercial applications": 4983,
      "standard benchmarks": 30870,
      "results compared": 28581,
      "stateoftheart opensource": 30967,
      "long sequence": 19176,
      "sequence modeling": 29604,
      "open llm": 23406,
      "nlp multimodal": 22940,
      "multimodal tasks": 22367,
      "despite successes": 7818,
      "llms high": 18671,
      "objective evaluations": 23204,
      "evaluations paper": 10037,
      "paper report": 24130,
      "solution significantly": 30480,
      "significantly reduce": 30080,
      "reduce llm": 27717,
      "llm training": 18374,
      "range evaluations": 27194,
      "evaluations existing": 10028,
      "existing evaluations": 10272,
      "evaluations focus": 10030,
      "minimize potential": 20197,
      "achieves performance": 1056,
      "layers improves": 17440,
      "improves factuality": 14377,
      "despite impressive": 7788,
      "content deviates": 5857,
      "seen pretraining": 29364,
      "decoding strategy": 7283,
      "reducing hallucinations": 27750,
      "llms does": 18545,
      "does require": 8536,
      "retrieved external": 28774,
      "additional finetuning": 1255,
      "later layers": 17414,
      "llms generally": 18633,
      "generation incorrect": 12521,
      "incorrect facts": 14584,
      "improves truthfulness": 14399,
      "tasks example": 32318,
      "performance llama": 24660,
      "llama family": 18100,
      "models truthfulqa": 22085,
      "primarily focus": 25910,
      "ask questions": 2554,
      "reliability trustworthiness": 27952,
      "accurately identify": 935,
      "provide reasonable": 26725,
      "investigate question": 15594,
      "different categories": 8056,
      "categories questions": 4226,
      "definitive answers": 7371,
      "tasks test": 32525,
      "tasks lack": 32386,
      "performance baseline": 24529,
      "research important": 28325,
      "gpt4 palm": 13108,
      "palm llama": 23991,
      "llama shown": 18143,
      "shown achieve": 29870,
      "llms ability": 18406,
      "llms pose": 18856,
      "pose challenges": 25158,
      "vietnamese language": 35272,
      "language limited": 16111,
      "leverage largescale": 17756,
      "instructionfollowing datasets": 15228,
      "datasets opensource": 7157,
      "opensource projects": 23541,
      "general domain": 12160,
      "specific medical": 30705,
      "medical domain": 19767,
      "instructional dataset": 15213,
      "utilize parameterefficient": 34961,
      "llms bloomz": 18452,
      "effectiveness methodology": 8959,
      "relevance accuracy": 27933,
      "responses evaluation": 28491,
      "evaluation process": 9994,
      "tasks analysis": 32243,
      "fake news": 10934,
      "news detection": 22880,
      "using finetuned": 34780,
      "paper considers": 24027,
      "considers possibility": 5728,
      "llama large": 18117,
      "detection finetuning": 7866,
      "finetuning peftlora": 11478,
      "peftlora based": 24440,
      "based approach": 3133,
      "used study": 34624,
      "study model": 31362,
      "finetuned following": 11312,
      "following tasks": 11704,
      "tasks analysing": 32241,
      "analysing text": 1909,
      "extracting named": 10759,
      "named entities": 22481,
      "sentiments obtained": 29578,
      "obtained results": 23257,
      "results finetuned": 28613,
      "finetuned llama": 11329,
      "reveal complex": 28792,
      "extracted sentiments": 10752,
      "sentiments named": 29574,
      "entities considered": 9630,
      "considered predictive": 5718,
      "predictive features": 25449,
      "features supervised": 11039,
      "supervised machine": 31684,
      "family large": 10975,
      "models lightweight": 21446,
      "opt family": 23591,
      "ranging 125m": 27217,
      "large collection": 16934,
      "collection diverse": 4929,
      "diverse data": 8422,
      "data time": 6893,
      "tokens current": 33221,
      "current input": 6492,
      "residual stream": 28392,
      "scale models": 29144,
      "textual data": 33024,
      "models operate": 21773,
      "power smaller": 25330,
      "smaller transformerbased": 30399,
      "10 million": 29,
      "million parameter": 20164,
      "produce coherent": 26140,
      "coherent english": 4892,
      "model python": 20740,
      "performance close": 24544,
      "use existing": 34528,
      "data way": 6911,
      "enhance learning": 9514,
      "learning process": 17634,
      "data follow": 6716,
      "approach focusing": 2286,
      "time common": 33110,
      "sense reasoning": 29508,
      "reasoning natural": 27429,
      "tasks comparable": 32268,
      "llms complex": 18478,
      "llms good": 18642,
      "step step": 31049,
      "including hallucinations": 14493,
      "biased generations": 3669,
      "data opensource": 6788,
      "attracted attention": 2752,
      "comprehensive capabilities": 5359,
      "network operations": 22700,
      "designed evaluating": 7732,
      "covering different": 6324,
      "available llms": 2988,
      "results gpt4": 28621,
      "gpt4 achieve": 13051,
      "open models": 23409,
      "like llama": 17882,
      "llama demonstrate": 18090,
      "significant potential": 30008,
      "integrated various": 15326,
      "sectors understanding": 29343,
      "crucial particularly": 6443,
      "autonomous driving": 2931,
      "framework investigate": 11872,
      "prominent llms": 26268,
      "including gpt35": 14485,
      "llms humans": 18681,
      "similarities llm": 30125,
      "llm human": 18320,
      "llms potential": 18858,
      "potential implications": 25262,
      "removing model": 28077,
      "targeted ablation": 32063,
      "models exhibit": 21225,
      "performance pretraining": 24719,
      "harm performance": 13438,
      "undesirable behaviors": 34291,
      "bad behavior": 3073,
      "given small": 12773,
      "generation minimal": 12549,
      "understanding interpreting": 34235,
      "need diverse": 22625,
      "zeroshot approach": 35952,
      "entities corpus": 9634,
      "gpt2 use": 12963,
      "news media": 22884,
      "model corpus": 20446,
      "previously encountered": 25897,
      "finetuning generate": 11409,
      "results encouraging": 28602,
      "astronomy large": 2672,
      "models excel": 21223,
      "highly specialized": 13668,
      "specialized domains": 30671,
      "domains like": 8628,
      "gap introduce": 12091,
      "finetuned llama2": 11334,
      "llama2 using": 18194,
      "adaptation model": 1186,
      "having significantly": 13474,
      "domainspecific model": 8654,
      "public release": 26843,
      "exploring large": 10618,
      "alignment work": 1784,
      "series flant5": 29636,
      "systems like": 31907,
      "framework prompt": 11891,
      "prompt design": 26318,
      "various recent": 35140,
      "study human": 31340,
      "human behavior": 13797,
      "behavior example": 3313,
      "semantically equivalent": 29484,
      "used measure": 34611,
      "llm prompts": 18355,
      "prompts lead": 26430,
      "lead different": 17463,
      "statistically significant": 31022,
      "large majority": 17228,
      "majority scenarios": 19449,
      "scores robust": 29281,
      "prompt template": 26348,
      "chatgpt llama2": 4487,
      "created humans": 6360,
      "decoding present": 7279,
      "novel inference": 23088,
      "accelerating large": 799,
      "approach characterized": 2248,
      "twostage process": 34046,
      "slightly lower": 30323,
      "intermediate layers": 15428,
      "original llm": 23711,
      "llm validate": 18379,
      "final output": 11179,
      "requires additional": 28247,
      "benchmarks llama2": 3459,
      "traditional chinese": 33343,
      "models comprehensive": 21078,
      "benchmark suite": 3414,
      "suite evaluation": 31599,
      "essential task": 9762,
      "field language": 11138,
      "generation language": 12531,
      "continue advance": 5986,
      "need effective": 22626,
      "benchmarks evaluate": 3440,
      "evaluate capabilities": 9820,
      "despite existence": 7775,
      "dataset address": 6938,
      "address gap": 1323,
      "leverage existing": 17746,
      "datasets tailored": 7179,
      "chinese benchmarks": 4541,
      "benchmarks encompass": 3439,
      "offer comprehensive": 23287,
      "framework enabling": 11846,
      "assessment language": 2616,
      "performance gpt35": 24619,
      "proprietary model": 26646,
      "highlight model": 13633,
      "comparable gpt35": 5079,
      "llms excel": 18574,
      "crafted prompts": 6341,
      "automate process": 2853,
      "process paper": 26077,
      "discrete prompt": 8280,
      "prompt optimization": 26338,
      "exhibit good": 10218,
      "discrete prompts": 8281,
      "language expressions": 16073,
      "approach allows": 2235,
      "processing capabilities": 26097,
      "llms efficient": 18553,
      "optimization performance": 23632,
      "prompts llms": 26431,
      "development set": 7975,
      "closed opensource": 4679,
      "datasets covering": 7086,
      "covering language": 6325,
      "tasks bigbench": 32254,
      "prompts existing": 26416,
      "automatic prompt": 2890,
      "prompt generation": 26327,
      "llms conventional": 18494,
      "better alpaca": 3591,
      "foundational large": 11805,
      "opendomain question": 23450,
      "applications like": 2166,
      "single language": 30209,
      "empirically analyze": 9246,
      "scenarios study": 29220,
      "multilingual data": 22304,
      "used tune": 34633,
      "tune llms": 33957,
      "par better": 24151,
      "findings serve": 11252,
      "language support": 16828,
      "good generating": 12820,
      "generating complex": 12415,
      "complex structured": 5295,
      "structured data": 31219,
      "power large": 25320,
      "structured outputs": 31226,
      "study assess": 31301,
      "assess capability": 2589,
      "current llms": 6511,
      "data propose": 6813,
      "propose structureaware": 26569,
      "improve ability": 14254,
      "perform comprehensive": 24479,
      "evaluation propose": 9996,
      "representative llms": 28184,
      "gptneox 20b": 13160,
      "carefully constructed": 4173,
      "datasets spanning": 7172,
      "tables based": 31987,
      "based analysis": 3132,
      "current model": 6515,
      "identify specific": 14016,
      "potential improvement": 25264,
      "address complex": 1318,
      "formatting requirements": 11761,
      "finetuning method": 11452,
      "based results": 3217,
      "reasoning comprehension": 27395,
      "weaknesses llms": 35466,
      "llms handling": 18666,
      "handling complex": 13414,
      "promising directions": 26287,
      "work code": 35677,
      "unlocking potential": 34413,
      "models dynamic": 21177,
      "widespread deployment": 35596,
      "enabling dynamic": 9319,
      "inference leveraging": 14789,
      "generative nlp": 12683,
      "making large": 19507,
      "approach boosts": 2243,
      "boosts model": 3828,
      "model efficiency": 20481,
      "need multiple": 22639,
      "various scenarios": 35142,
      "unlock power": 34411,
      "layers transformers": 17448,
      "transformers generating": 33782,
      "target output": 32057,
      "components original": 5316,
      "model minimizing": 20642,
      "storage requirements": 31078,
      "method demonstrated": 19897,
      "tune llama": 33956,
      "llama 13b": 18059,
      "dataset instruction": 7004,
      "comparison standard": 5200,
      "usage inference": 34504,
      "contrastive decoding": 6021,
      "improves reasoning": 14392,
      "demonstrate contrastive": 7443,
      "method proposed": 19960,
      "li et": 17808,
      "greedy decoding": 13281,
      "perceived quality": 24457,
      "generation contrastive": 12480,
      "difference likelihood": 8035,
      "palm 2l": 23989,
      "gsm8k math": 13327,
      "math word": 19675,
      "benchmark addition": 3352,
      "analysis suggests": 1969,
      "improves existing": 14376,
      "abstract reasoning": 773,
      "reasoning errors": 27407,
      "making powerful": 19514,
      "powerful general": 25337,
      "general purpose": 12184,
      "purpose method": 26883,
      "method generating": 19926,
      "typically require": 34081,
      "large gpu": 16953,
      "gpu memory": 13173,
      "reduce gpu": 27711,
      "accuracy existing": 877,
      "solutions provide": 30485,
      "tensor core": 32722,
      "based key": 3182,
      "main bottleneck": 19387,
      "bottleneck generative": 3839,
      "matrix multiplications": 19694,
      "propose general": 26516,
      "matrix multiplication": 19693,
      "memory bandwidth": 19802,
      "bandwidth bottleneck": 3089,
      "endtoend performance": 9438,
      "effective software": 8895,
      "sparse data": 30611,
      "data extraction": 6706,
      "significantly lower": 30068,
      "inference cost": 14770,
      "examples natural": 10137,
      "reducing need": 27757,
      "need extensive": 22628,
      "limited capability": 17942,
      "report present": 28122,
      "largescale multilingual": 17365,
      "models containing": 21103,
      "matches outperforms": 19653,
      "public benchmarks": 26832,
      "like mmlu": 17888,
      "mmlu cmmlu": 20315,
      "medicine law": 19773,
      "pretraining model": 25821,
      "benefit research": 3483,
      "store information": 31080,
      "evaluating faithfulness": 9895,
      "help address": 13502,
      "address develop": 1322,
      "modes evaluation": 22176,
      "evaluation natural": 9979,
      "individual neurons": 14718,
      "text input": 32898,
      "apply framework": 2207,
      "gpt2 xl": 12969,
      "high error": 13566,
      "error rates": 9716,
      "paper critically": 24029,
      "modern llms": 22165,
      "interactive llms": 15395,
      "gpt4 bard": 13058,
      "bard llama": 3098,
      "careful attention": 4168,
      "systems built": 31887,
      "practical terms": 25373,
      "amounts compute": 1883,
      "resources does": 28433,
      "llms llms": 18781,
      "architectures incorporate": 2465,
      "superior capabilities": 31645,
      "processing understanding": 26133,
      "understanding language": 34237,
      "language applications": 16042,
      "remain underexplored": 27987,
      "crucial step": 6447,
      "deeper understanding": 7349,
      "related concepts": 27854,
      "effective solution": 8896,
      "task automated": 32081,
      "explanation generation": 10536,
      "generation present": 12573,
      "present evaluate": 25528,
      "evaluate framework": 9838,
      "framework called": 11834,
      "questions large": 27117,
      "evaluation model": 9977,
      "model framework": 20538,
      "framework generates": 11861,
      "generates highquality": 12402,
      "instruction prompt": 15174,
      "llama213b gpt4": 18198,
      "gpt4 generate": 13080,
      "higher quality": 13603,
      "quality explanations": 26959,
      "written students": 35864,
      "datasets findings": 7117,
      "promising path": 26292,
      "models educational": 21179,
      "various data": 35084,
      "627b tokens": 466,
      "tokens extensive": 33226,
      "fundamental characteristics": 11974,
      "different sources": 8141,
      "local single": 19135,
      "single source": 30222,
      "affect performance": 1539,
      "performance trained": 24787,
      "slimpajama dataset": 30325,
      "using 13b": 34723,
      "best configuration": 3556,
      "configuration outperforms": 5655,
      "13b model": 127,
      "number training": 23169,
      "tokens significant": 33245,
      "significant margin": 29994,
      "13b models": 129,
      "small language": 30346,
      "text style": 32951,
      "1000 sentences": 44,
      "impact large": 14127,
      "llm shown": 18366,
      "shown promise": 29906,
      "privacy concerns": 25951,
      "high deployment": 13564,
      "deployment costs": 7646,
      "explore effectiveness": 10582,
      "learning propose": 17637,
      "automated evaluation": 2860,
      "evaluation text": 10020,
      "quality based": 26944,
      "compare approach": 5104,
      "methods model": 20067,
      "models effective": 21180,
      "field machine": 11139,
      "translation recent": 33848,
      "nmt systems": 22980,
      "fail handle": 10902,
      "cases large": 4200,
      "llms emerged": 18555,
      "emerged promising": 9164,
      "promising alternative": 26281,
      "performance traditional": 24786,
      "nmt models": 22979,
      "models introducing": 21405,
      "outputs paper": 23896,
      "word senses": 35648,
      "match outperform": 19644,
      "outperform stateoftheart": 23787,
      "nllb language": 22915,
      "language directions": 16065,
      "provides valuable": 26767,
      "adapting llms": 1212,
      "llms better": 18448,
      "cultural characteristics": 6460,
      "current mainstream": 6513,
      "mainstream models": 19412,
      "models significant": 21973,
      "significant concerns": 29972,
      "address paper": 1348,
      "proposes comprehensive": 26626,
      "texts supervised": 33002,
      "gpt4 responses": 13113,
      "ai feedback": 1608,
      "reward model": 28859,
      "evaluations reveal": 10039,
      "model dubbed": 20476,
      "sets stateoftheart": 29720,
      "benchmarks including": 3449,
      "newly introduced": 22871,
      "value alignment": 35023,
      "evaluated gpt4": 9876,
      "limited scale": 17963,
      "codes data": 4851,
      "llms represent": 18910,
      "interact computers": 15380,
      "dedicated hardware": 7304,
      "execution study": 10200,
      "study evaluate": 31326,
      "based 13": 3131,
      "llama models": 18132,
      "quantization process": 27011,
      "models considered": 21097,
      "koala vicuna": 15940,
      "models developed": 21152,
      "analysis revealed": 1957,
      "achieved accuracy": 999,
      "original texts": 23722,
      "english translations": 9494,
      "models measuring": 21712,
      "time required": 33140,
      "average 13": 3008,
      "approximately 20": 2414,
      "process queries": 26078,
      "calculations large": 3968,
      "models highquality": 21359,
      "conversational datasets": 6099,
      "successful development": 31533,
      "development intelligent": 7956,
      "systems utilize": 31929,
      "generated using": 12398,
      "using advanced": 34727,
      "gpt4 models": 13100,
      "models common": 21065,
      "common strategy": 5015,
      "creating datasets": 6365,
      "pose challenge": 25157,
      "capabilities limitations": 4037,
      "mathematical reasoning": 19682,
      "limitation introduce": 17912,
      "uses python": 34714,
      "approach notably": 2318,
      "quality synthetic": 26980,
      "synthetic conversation": 31848,
      "datasets especially": 7104,
      "expert evaluations": 10509,
      "model effectively": 20480,
      "reversal curse": 28820,
      "autoregressive large": 2947,
      "trained sentence": 33425,
      "instance model": 15110,
      "correct answer": 6191,
      "curse finetuning": 6551,
      "finetuning gpt3": 11413,
      "chatgpt gpt35": 4475,
      "questions like": 27120,
      "insights dataset": 15071,
      "instructiontuned lms": 15297,
      "lms led": 19095,
      "usage development": 34502,
      "detailed information": 7839,
      "automate model": 2852,
      "generation introduce": 12527,
      "questionanswer pairs": 27080,
      "models cover": 21113,
      "aspects model": 2578,
      "training configurations": 33456,
      "architecture details": 2442,
      "details training": 7847,
      "capabilities lms": 4042,
      "lms generating": 19087,
      "initial experiments": 14959,
      "llama galactica": 18104,
      "showcase significant": 29840,
      "gap understanding": 12114,
      "textual responses": 33037,
      "models automate": 20996,
      "automate generation": 2851,
      "process complete": 26053,
      "complete dataset": 5249,
      "dataset available": 6941,
      "impressive skills": 14249,
      "tasks llms": 32408,
      "llms evaluated": 18571,
      "evaluated various": 9885,
      "thorough evaluation": 33072,
      "llms benchmark": 18446,
      "benchmark performance": 3405,
      "performance bengali": 24532,
      "classification sentiment": 4611,
      "zeroshot llms": 35984,
      "current sota": 6529,
      "models tasks": 22047,
      "performance opensource": 24703,
      "efforts develop": 9088,
      "develop better": 7911,
      "llms large": 18743,
      "enormous parameter": 9590,
      "extremely high": 10797,
      "compute power": 5497,
      "challenges practical": 4369,
      "practical deployment": 25363,
      "models distillation": 21161,
      "studies explore": 31267,
      "explore potential": 10597,
      "leveraging llms": 17788,
      "work focusing": 35715,
      "table reasoning": 31984,
      "reasoning skills": 27452,
      "specifically tailored": 30757,
      "scientific tabletotext": 29257,
      "reasoning distillation": 27405,
      "distillation approach": 8336,
      "distilling llms": 8357,
      "llms tailored": 18989,
      "models experimental": 21235,
      "results shown": 28682,
      "distilled data": 8351,
      "data achieves": 6591,
      "improvement compared": 14336,
      "traditionally finetuned": 33355,
      "specific llms": 30704,
      "generation dataset": 12482,
      "test study": 32789,
      "ethical considerations": 9805,
      "based relevance": 3216,
      "study shows": 31399,
      "gpt3 exhibit": 12992,
      "better random": 3621,
      "chatgpt llama2chat": 4489,
      "palm2 gpt4": 23994,
      "gpt4 significantly": 13118,
      "observe models": 23232,
      "perform consistently": 24481,
      "gaps understanding": 12118,
      "rank decomposition": 27223,
      "code llms": 4777,
      "parameters model": 24273,
      "speedup modern": 30801,
      "hardware unlike": 13433,
      "linear layers": 17989,
      "efficient kernels": 9041,
      "floating point": 11622,
      "study potential": 31374,
      "layers models": 17443,
      "pass1 score": 24376,
      "single a100": 30196,
      "quantization method": 27009,
      "compression gains": 5415,
      "model reduces": 20746,
      "reduces memory": 27736,
      "similar gains": 30102,
      "tuning work": 34020,
      "work shows": 35786,
      "promising new": 26290,
      "llm compression": 18283,
      "use knowledge": 34539,
      "logical reasoning": 19157,
      "reasoning remains": 27448,
      "stored knowledge": 31082,
      "knowledge retrieval": 15904,
      "chain thoughts": 4293,
      "perform poorly": 24500,
      "primary contribution": 25920,
      "synthetic dataset": 31853,
      "inherent weaknesses": 14952,
      "weaknesses language": 35465,
      "model efficiently": 20484,
      "data knowledge": 6745,
      "paradigm efficient": 24155,
      "domainspecific text": 8658,
      "faces challenge": 10828,
      "data high": 6726,
      "methods large": 20057,
      "llms gained": 18626,
      "models slms": 21987,
      "1b parameters": 198,
      "offer significant": 23293,
      "domainspecific tasks": 8657,
      "tasks given": 32349,
      "investigate potential": 15593,
      "potential slms": 25298,
      "promptbased model": 26364,
      "220m parameters": 252,
      "parameters achieve": 24221,
      "approximately 75": 2415,
      "limited labeled": 17952,
      "data shows": 6862,
      "validate effectiveness": 34993,
      "pivotal observation": 24979,
      "designed prompts": 7740,
      "accuracy exceeding": 876,
      "optimized prompt": 23647,
      "findings underscore": 11259,
      "underscore promise": 34177,
      "models solving": 21991,
      "nlp problems": 22942,
      "recent developments": 27515,
      "promise enhancing": 26274,
      "enhancing capabilities": 9557,
      "nlp despite": 22930,
      "problemsolving abilities": 26038,
      "llms gap": 18627,
      "present unique": 25560,
      "benchmarking dataset": 3422,
      "questions spanning": 27132,
      "spanning various": 30600,
      "questions context": 27102,
      "context multiple": 5903,
      "information diverse": 14862,
      "question types": 27075,
      "answer math": 2046,
      "advanced prompting": 1440,
      "strategies like": 31109,
      "effectiveness advanced": 8938,
      "especially smaller": 9747,
      "like llama2": 17886,
      "llama2 13b": 18159,
      "problemsolving skills": 26041,
      "comprehension llms": 5344,
      "improvements natural": 14359,
      "coding tasks": 4869,
      "tasks ability": 32228,
      "formal languages": 11745,
      "proficiency various": 26184,
      "various llms": 35113,
      "llms created": 18497,
      "set tasks": 29708,
      "tasks probe": 32458,
      "ability parse": 709,
      "understand analyze": 34188,
      "create knowledge": 6351,
      "tasks embodying": 32306,
      "integrated automated": 15324,
      "gpt4 claude": 13063,
      "freely accessible": 11915,
      "analysis offers": 1946,
      "indepth understanding": 14679,
      "understanding strengths": 34272,
      "commercial models": 4990,
      "models fall": 21260,
      "crucial requirement": 6445,
      "thoughts prompting": 33083,
      "extensive text": 10714,
      "textrelated tasks": 32987,
      "encounter challenges": 9392,
      "proposed means": 26604,
      "enhance llms": 9517,
      "solving math": 30510,
      "word problems": 35645,
      "primary aim": 25915,
      "aim research": 1649,
      "skills using": 30316,
      "cot prompting": 6281,
      "research provide": 28351,
      "following contributions": 11689,
      "essays dataset": 9754,
      "dataset previously": 7023,
      "use cot": 34522,
      "approach training": 2349,
      "suggest models": 31575,
      "models llama7b": 21467,
      "cohen kappa": 4885,
      "kappa score": 15740,
      "important note": 14206,
      "user privacy": 34665,
      "allowing users": 1808,
      "metric text": 20126,
      "wide spectrum": 35565,
      "tasks different": 32297,
      "evaluation methods": 9973,
      "methods provide": 20082,
      "guided natural": 13353,
      "language instruction": 16097,
      "based llama2": 3192,
      "meticulously curated": 20116,
      "datasets dataset": 7090,
      "analysis collected": 1912,
      "outputs large": 23894,
      "large variety": 17286,
      "variety models": 35064,
      "quantitatively assess": 26997,
      "correlation human": 6219,
      "human ratings": 13859,
      "surpass best": 31733,
      "best existing": 3558,
      "possibility building": 25207,
      "testing limits": 32805,
      "llm pretraining": 18350,
      "pretraining diverse": 25793,
      "present web": 25562,
      "web pages": 35477,
      "semistructured data": 29503,
      "available internet": 2984,
      "approach large": 2305,
      "used solve": 34623,
      "solve diverse": 30490,
      "classification problems": 4605,
      "models specialized": 21997,
      "specialized task": 30675,
      "model works": 20872,
      "tasks significant": 32503,
      "significant degradation": 29976,
      "pretraining stage": 25840,
      "t5 data": 31939,
      "surrounding context": 31771,
      "context downstream": 5886,
      "selfsupervised objectives": 29439,
      "significantly boost": 30037,
      "specific pretraining": 30710,
      "sequence models": 29605,
      "finetuned variants": 11360,
      "models reasoning": 21899,
      "essential understanding": 9763,
      "limited scope": 17964,
      "consistent evaluations": 5737,
      "different studies": 8144,
      "datasets encompassing": 7102,
      "encompassing various": 9390,
      "temporal aspects": 32694,
      "facilitate comprehensive": 10835,
      "llms conduct": 18484,
      "evaluation using": 10024,
      "learning scenarios": 17650,
      "scenarios additionally": 29200,
      "additionally employ": 1282,
      "bertbased models": 3541,
      "establish baseline": 9766,
      "findings indicate": 11238,
      "indicate models": 14690,
      "models trail": 22063,
      "spur progress": 30833,
      "llms future": 18623,
      "providing nuanced": 26780,
      "nuanced understanding": 23133,
      "data recent": 6826,
      "advancements llms": 1469,
      "demonstrated potential": 7539,
      "focus tasks": 11658,
      "simple reasoning": 30162,
      "reasoning processes": 27439,
      "gap remains": 12108,
      "reasoning events": 27408,
      "limitation existing": 17911,
      "introduce task": 15535,
      "requires multiple": 28260,
      "multiple reasoning": 22414,
      "reasoning multiple": 27428,
      "multiple events": 22390,
      "task offers": 32170,
      "ai applications": 1597,
      "applications support": 2178,
      "dataset explainable": 6988,
      "reasoning paths": 27431,
      "paths using": 24402,
      "using novel": 34861,
      "based dataset": 3150,
      "propose opensource": 26558,
      "llm series": 18365,
      "based foundation": 3162,
      "ability instruction": 685,
      "performance method": 24675,
      "prediction explanation": 25423,
      "aim provide": 1647,
      "explore large": 10587,
      "learning architecture": 17539,
      "zeroshot prompting": 35990,
      "automatically using": 2923,
      "accuracy using": 914,
      "greater accuracy": 13267,
      "task best": 32086,
      "models combined": 21064,
      "automated methods": 2869,
      "models outputs": 21786,
      "feedback utterances": 11074,
      "negatively correlated": 22665,
      "gained significant": 12065,
      "academia industry": 783,
      "fewshot generalization": 11104,
      "capabilities opensource": 4057,
      "llms finetuning": 18611,
      "realworld tasks": 27349,
      "tasks generating": 32347,
      "generating responses": 12442,
      "token classification": 33185,
      "tasks classification": 32262,
      "label space": 15952,
      "prior research": 25938,
      "llms outperform": 18834,
      "bert prompting": 3526,
      "latent representations": 17409,
      "representations llms": 28168,
      "adaptation llms": 1183,
      "llms aims": 18428,
      "llm finetuned": 18306,
      "representations final": 28160,
      "loss model": 19249,
      "minimize loss": 20196,
      "consistent improvements": 5739,
      "baselines like": 3268,
      "benchmarking llms": 3427,
      "recently showcased": 27622,
      "showcased remarkable": 29844,
      "domains clinical": 8613,
      "clinical medicine": 4661,
      "diverse user": 8471,
      "evaluating diverse": 9892,
      "aspects llms": 2577,
      "abilities study": 643,
      "textdavinci003 gpt35turbo": 32981,
      "gpt4 llama27b": 13093,
      "llama27b llama213b": 18206,
      "bypass safety": 3948,
      "openly accessible": 23473,
      "games large": 12080,
      "llms effective": 18551,
      "incorrect outputs": 14586,
      "underscores need": 34181,
      "need development": 22623,
      "intelligent agents": 15364,
      "agents capable": 1567,
      "capability requires": 4102,
      "requires complex": 28249,
      "state tracking": 30911,
      "tracking reasoning": 33328,
      "paper offer": 24081,
      "models systematically": 22039,
      "evaluate various": 9869,
      "significant differences": 29978,
      "differences performance": 8040,
      "model finally": 20521,
      "enhance reasoning": 9528,
      "vicuna models": 35254,
      "lead significant": 17470,
      "improvement hope": 14340,
      "offers insights": 23310,
      "autonomous agents": 2930,
      "september 2023": 29587,
      "generating validating": 12447,
      "measuring consistency": 19743,
      "lms propose": 19105,
      "propose finetune": 26513,
      "math questions": 19671,
      "qa instruction": 26910,
      "llms fall": 18601,
      "domain experts": 8562,
      "intricate nature": 15484,
      "issues introduce": 15668,
      "domain instruction": 8568,
      "based multiagent": 3198,
      "multiagent collaboration": 22266,
      "additionally construct": 1277,
      "higher level": 13598,
      "tasks gains": 32344,
      "intelligence capabilities": 15354,
      "soon available": 30523,
      "models represent": 21928,
      "llms sparked": 18961,
      "systems just": 31903,
      "reflect real": 27792,
      "real world": 27313,
      "learned representations": 17522,
      "llama2 family": 18176,
      "models discover": 21159,
      "discover llms": 8267,
      "llms learn": 18749,
      "linear representations": 17996,
      "multiple scales": 22417,
      "representations robust": 28174,
      "robust prompting": 28939,
      "entity types": 9655,
      "spatial temporal": 30642,
      "decoderonly language": 7255,
      "models standard": 22009,
      "scale poorly": 29147,
      "propose solution": 26568,
      "solution based": 30470,
      "method models": 19947,
      "offtheshelf models": 23330,
      "experiments language": 10454,
      "modeling question": 20905,
      "drastically reducing": 8722,
      "terms time": 32751,
      "compression ratio": 5425,
      "achieving nearly": 1097,
      "costs large": 6269,
      "llms exploded": 18589,
      "exploded popularity": 10553,
      "prior stateoftheart": 25939,
      "technologies increasingly": 32677,
      "domains law": 8627,
      "finance medicine": 11214,
      "medicine models": 19774,
      "llms despite": 18532,
      "despite large": 7793,
      "models called": 21037,
      "llms increasing": 18710,
      "deployment various": 7660,
      "paper experiments": 24044,
      "conducted study": 5640,
      "inference llms": 14790,
      "benchmark conduct": 3363,
      "conduct preliminary": 5615,
      "inference performance": 14796,
      "llama recent": 18140,
      "recent stateoftheart": 27550,
      "stateoftheart llm": 30944,
      "llm developed": 18288,
      "developed meta": 7929,
      "meta ai": 19857,
      "gpus nvidia": 13179,
      "datasets alpaca": 7063,
      "research practice": 28344,
      "present results": 25551,
      "multigpu inference": 22282,
      "inference using": 14822,
      "study llm": 31358,
      "implicit representations": 14177,
      "representations knowledge": 28161,
      "remains open": 28005,
      "models contain": 21102,
      "contain various": 5830,
      "responsible encoding": 28522,
      "knowledge model": 15881,
      "knowledge models": 15882,
      "models minimizing": 21725,
      "adverse effects": 1525,
      "original language": 23710,
      "demonstrate method": 7473,
      "code debugging": 4745,
      "solving problem": 30515,
      "solution directly": 30474,
      "conversational agents": 6093,
      "human instruction": 13825,
      "evaluation paper": 9984,
      "manually created": 19565,
      "zeroshot chain": 35960,
      "chain thought": 4290,
      "thought prompting": 33081,
      "gpt4 code": 13066,
      "code datasets": 4744,
      "freely available": 11916,
      "models good": 21322,
      "success wide": 31529,
      "spectrum tasks": 30774,
      "tasks face": 32327,
      "face limitations": 10824,
      "spatial reasoning": 30640,
      "benchmark termed": 3417,
      "benchmark evaluates": 3384,
      "evaluates llms": 9888,
      "path planning": 24398,
      "planning tasks": 25003,
      "constraints leveraging": 5791,
      "systematically investigate": 31882,
      "investigate llms": 15587,
      "gpt4 different": 13072,
      "finetuning experimental": 11400,
      "results promise": 28658,
      "fewshot gpt4": 11106,
      "reason act": 27355,
      "model calls": 20407,
      "solve complex": 30489,
      "unfortunately existing": 34313,
      "prompt templates": 26349,
      "systematic approach": 31867,
      "approach developing": 2260,
      "computational graphs": 5468,
      "prompting finetuning": 26373,
      "reasoning techniques": 27459,
      "conduct case": 5585,
      "studies showing": 31282,
      "retrieval answer": 28731,
      "answer complex": 2041,
      "outperform standard": 23786,
      "small lms": 30353,
      "lms like": 19096,
      "competitive approaches": 5219,
      "proprietary gpt35": 26636,
      "opportunities challenges": 23581,
      "despite rapid": 7805,
      "rapid advancements": 27246,
      "advancements artificial": 1460,
      "recently emergence": 27596,
      "rapid adoption": 27238,
      "advanced large": 1425,
      "study investigating": 31349,
      "ai genai": 1609,
      "gap researchers": 12109,
      "researchers practitioners": 28379,
      "bridging gap": 3874,
      "unprecedented capabilities": 34419,
      "capabilities generate": 4020,
      "content based": 5852,
      "based learning": 3190,
      "questions future": 27110,
      "study delves": 31313,
      "perception using": 24465,
      "implementation framework": 14162,
      "framework provides": 11892,
      "subsequent research": 31446,
      "llm prompting": 18354,
      "prompting prompt": 26392,
      "llms poorly": 18854,
      "class discrete": 4583,
      "dynamical systems": 8765,
      "explore prompt": 10598,
      "theory investigate": 33059,
      "output token": 23883,
      "control input": 6052,
      "llm output": 18342,
      "parameter matrices": 24189,
      "matrices present": 19691,
      "short prompt": 29814,
      "prompt sequences": 26343,
      "analysis llms": 1939,
      "llms demonstrates": 18529,
      "input sequences": 15029,
      "enhancing language": 9561,
      "attention head": 2718,
      "single attention": 30198,
      "overall model": 23908,
      "model calibration": 20405,
      "component language": 5309,
      "components results": 5317,
      "task interactive": 32143,
      "pretrained largescale": 25702,
      "abilities realworld": 637,
      "realworld knowledge": 27341,
      "llm capabilities": 18279,
      "capabilities english": 4013,
      "increasingly vital": 14650,
      "datasets work": 7191,
      "multitask language": 22449,
      "understanding benchmark": 34211,
      "empirical evaluations": 9222,
      "models bloomz": 21033,
      "falcon perform": 10941,
      "joint entity": 15712,
      "extraction process": 10772,
      "entity pairs": 9647,
      "using single": 34911,
      "focus problem": 11655,
      "problem training": 26016,
      "entity mentions": 9646,
      "corresponding entity": 6226,
      "knowledge base": 15816,
      "key challenge": 15755,
      "effectiveness supervised": 8968,
      "supervised learning": 31682,
      "learning applications": 17535,
      "applications existing": 2152,
      "research primarily": 28346,
      "limiting effectiveness": 17977,
      "learning framework": 17571,
      "includes new": 14452,
      "experiments datasets": 10430,
      "extraction performance": 10771,
      "benchmarking large": 3423,
      "models augmented": 20995,
      "task natural": 32164,
      "need adapt": 22618,
      "dataset tailored": 7044,
      "llms employing": 18559,
      "output formats": 23867,
      "extensive evaluations": 10687,
      "t5 flant5": 31946,
      "generalizing unseen": 12238,
      "chatgpt exhibits": 4467,
      "performance solely": 24758,
      "techniques work": 32672,
      "work paves": 35744,
      "utilization llms": 34957,
      "llms information": 18720,
      "emerged pivotal": 9161,
      "solutions data": 30484,
      "data imbalance": 6730,
      "topic models": 33286,
      "data level": 6756,
      "data poses": 6799,
      "poses unique": 25176,
      "unique challenges": 34356,
      "hierarchical structure": 13544,
      "efficacy generated": 8987,
      "data demonstrating": 6674,
      "effectively address": 8909,
      "address aforementioned": 1307,
      "generate high": 12282,
      "scientific text": 29259,
      "help model": 13510,
      "embeddings improve": 9141,
      "finetuning improved": 11417,
      "embedding vectors": 9134,
      "finetuning llama27b": 11445,
      "llama27b using": 18211,
      "using alpaca": 34731,
      "using noisy": 34860,
      "datasets models": 7149,
      "10 improvement": 28,
      "models application": 20982,
      "application machine": 2135,
      "increasingly common": 14633,
      "predictive power": 25454,
      "spatial information": 30638,
      "key indicators": 15771,
      "method effectively": 19908,
      "performance measured": 24673,
      "use information": 34537,
      "information directly": 14861,
      "directly prompt": 8241,
      "observe gpt35": 23228,
      "gpt35 outperforms": 13031,
      "outperforms llama": 23829,
      "dataset experiments": 6987,
      "reveal llms": 28803,
      "llms remarkably": 18909,
      "limitations existing": 17919,
      "remarkable capabilities": 28031,
      "domains including": 8621,
      "achieve similar": 984,
      "similar better": 30098,
      "present comprehensive": 25520,
      "llama gpt": 18109,
      "retrievalaugmented generation": 28761,
      "techniques combine": 32630,
      "retrieval generation": 28741,
      "generation capabilities": 12466,
      "capabilities prompting": 4064,
      "strategies improve": 31107,
      "improve llms": 14274,
      "performance demonstrate": 24565,
      "analysis highlights": 1925,
      "gpt4s ability": 13133,
      "ability achieve": 652,
      "passing score": 24385,
      "generalpurpose models": 12256,
      "accuracy experiments": 878,
      "experiments gpt4": 10447,
      "compared human": 5140,
      "performance suggests": 24773,
      "explore models": 10593,
      "capacity address": 4125,
      "address general": 1328,
      "contribute meaningfully": 6030,
      "continual learning": 5974,
      "models aligned": 20972,
      "demonstrate exceptional": 7454,
      "ensuring safety": 9612,
      "aligned llms": 1740,
      "overlooked existing": 23947,
      "learning benchmarks": 17545,
      "lack sufficient": 16003,
      "models potential": 21826,
      "tuning paper": 34001,
      "novel benchmark": 23064,
      "benchmark designed": 3378,
      "designed evaluate": 7731,
      "consists distinct": 5763,
      "distinct datasets": 8364,
      "capabilities code": 4005,
      "unified format": 34327,
      "llms experiments": 18588,
      "exhibit significant": 10227,
      "general ability": 12156,
      "example accuracy": 10109,
      "llama2chat 13b": 18216,
      "challenge finding": 4313,
      "finding suitable": 11227,
      "achieving performance": 1099,
      "preserving original": 25611,
      "prowess llms": 26791,
      "llms empirical": 18557,
      "empirical findings": 9226,
      "tasks inherently": 32373,
      "motivated introduce": 22228,
      "effectively reducing": 8930,
      "sparse finetuning": 30612,
      "acceleration large": 804,
      "models consider": 21095,
      "llms specialized": 18962,
      "specialized tasks": 30676,
      "especially high": 9737,
      "detailed study": 7842,
      "enables accurate": 9293,
      "model types": 20846,
      "sparse llms": 30616,
      "cpu gpu": 6336,
      "standard approach": 30868,
      "reducing memory": 27754,
      "accuracy t5": 913,
      "language translation": 16837,
      "speech translation": 30790,
      "generation time": 12622,
      "accuracy drops": 872,
      "gpu inference": 13171,
      "compatible quantization": 5205,
      "instructiontuning llms": 15307,
      "paper makes": 24080,
      "indepth empirical": 14676,
      "systematically explore": 31881,
      "explore impact": 10586,
      "impact llm": 14130,
      "parameterefficient methods": 24213,
      "methods instruction": 20051,
      "conduct experiment": 5601,
      "open chinese": 23389,
      "paper release": 24129,
      "release powerful": 27917,
      "llms comparable": 18474,
      "opensource alternatives": 23485,
      "performance address": 24519,
      "metric performance": 20123,
      "optimal model": 23614,
      "task considering": 32099,
      "performance cost": 24559,
      "source models": 30568,
      "models varying": 22111,
      "varying sizes": 35179,
      "sizes 7b": 30293,
      "small memory": 30357,
      "memory footprints": 19816,
      "vicuna benchmark": 35247,
      "leverage llms": 17757,
      "compromising performance": 5441,
      "reducing costs": 27745,
      "democratizing access": 7427,
      "models zeroshot": 22143,
      "time series": 33141,
      "series forecasting": 29638,
      "nexttoken prediction": 22894,
      "series models": 29641,
      "trained downstream": 33394,
      "performance propose": 24723,
      "distributions tokens": 8403,
      "success llms": 31518,
      "llms time": 19004,
      "naturally represent": 22586,
      "salient features": 29061,
      "textual information": 33030,
      "information answer": 14855,
      "questions help": 27113,
      "explain predictions": 10524,
      "size generally": 30250,
      "generally improves": 12242,
      "performance time": 24785,
      "uncertainty calibration": 34116,
      "result alignment": 28539,
      "text entry": 32853,
      "process making": 26073,
      "effective efficient": 8867,
      "deep learningbased": 7334,
      "collection model": 4931,
      "learning capability": 17548,
      "capability large": 4093,
      "models gpt35": 21329,
      "unique feature": 34360,
      "feature allows": 11021,
      "various text": 35153,
      "prompting gpt35": 26375,
      "gpt35 model": 13030,
      "models specialize": 21996,
      "specific text": 30720,
      "expertise prompt": 10514,
      "address introduce": 1330,
      "agent designed": 1559,
      "complex prompts": 5284,
      "meet specific": 19781,
      "specific needs": 30708,
      "conducted user": 5641,
      "study involving": 31350,
      "prompts results": 26439,
      "increase similarity": 14605,
      "assessing capabilities": 2610,
      "capabilities human": 4024,
      "structure llms": 31214,
      "paper based": 24017,
      "meticulously annotated": 20114,
      "human test": 13869,
      "test dataset": 32764,
      "dataset based": 6942,
      "structures llms": 31231,
      "llms gain": 18625,
      "cognitive capabilities": 4878,
      "emphasizes significance": 9214,
      "investigating llms": 15610,
      "patterns llms": 24412,
      "llms shedding": 18934,
      "researchers advance": 28371,
      "development utilization": 7977,
      "finegrained evaluation": 11272,
      "capability language": 4091,
      "llm gpt4": 18318,
      "longform responses": 19207,
      "using proprietary": 34888,
      "closedsource nature": 4690,
      "costs work": 6277,
      "language feedback": 16076,
      "generated gpt4": 12362,
      "gpt4 using": 13126,
      "llm assess": 18271,
      "assess given": 2594,
      "provided user": 26743,
      "human evaluators": 13812,
      "par gpt4": 24152,
      "greatly outperforms": 13274,
      "compared opensourced": 5156,
      "reward models": 28861,
      "models explicitly": 21238,
      "explicitly trained": 10551,
      "preference datasets": 25468,
      "highlighting potential": 13647,
      "model opensource": 20664,
      "code dataset": 4742,
      "models materials": 21704,
      "materials science": 19665,
      "data curation": 6666,
      "model targeted": 20824,
      "helps alleviate": 13523,
      "available open": 2992,
      "billionparameter language": 3724,
      "model specialized": 20804,
      "data prompting": 6811,
      "available large": 2985,
      "models generation": 21314,
      "construct dataset": 5795,
      "dataset multiple": 7014,
      "measure quality": 19736,
      "quality dataset": 26950,
      "multiple dimensions": 22386,
      "dimensions including": 8200,
      "iteratively generate": 15690,
      "progressively better": 26234,
      "existing language": 10281,
      "evaluation analyze": 9919,
      "studies understand": 31290,
      "limitations code": 17916,
      "code relevant": 4808,
      "datasets publicly": 7165,
      "applied question": 2196,
      "answering text": 2078,
      "rank set": 27225,
      "set candidate": 29674,
      "different predictions": 8122,
      "model decoding": 20454,
      "decoding approach": 7271,
      "language sentences": 16821,
      "develop computational": 7912,
      "applied large": 2191,
      "reasoning mathematical": 27421,
      "mathematical problemsolving": 19681,
      "benchmarks observe": 3465,
      "highlight promise": 13636,
      "fundamental challenges": 11973,
      "tuning human": 33981,
      "understanding human": 34230,
      "research proposes": 28349,
      "learning methodology": 17603,
      "nature human": 22589,
      "key steps": 15785,
      "synthetic instruction": 31857,
      "pipeline designed": 24966,
      "instruction specifically": 15178,
      "generation framework": 12507,
      "taxonomy classic": 32576,
      "questions presented": 27126,
      "utilizing information": 34972,
      "yields significant": 35927,
      "performance enhancements": 24582,
      "enhancements compared": 9544,
      "approach consistently": 2252,
      "benchmarks hope": 3445,
      "similarity human": 30129,
      "learning rank": 17639,
      "dataset recent": 7030,
      "recent pretrained": 27540,
      "limited range": 17960,
      "alleviate issue": 1789,
      "lack supervision": 16004,
      "unsupervised approaches": 34454,
      "instead propose": 15124,
      "synthetic context": 31847,
      "context retrieval": 5915,
      "retrieval training": 28757,
      "llm using": 18378,
      "based bert": 3138,
      "ner task": 22680,
      "models sensitivity": 21963,
      "spurious features": 30836,
      "features prompt": 11037,
      "llms adopted": 18427,
      "fundamental component": 11976,
      "crucial accurately": 6437,
      "choices prompt": 4557,
      "influence model": 14838,
      "effectively using": 8934,
      "modern pretrained": 22169,
      "work focus": 35711,
      "llms extremely": 18597,
      "accuracy points": 901,
      "using llama213b": 34822,
      "tuning analysis": 33966,
      "evaluating llms": 9905,
      "performance single": 24754,
      "fixed prompt": 11579,
      "prompt format": 26325,
      "systematic analysis": 31866,
      "analysis propose": 1950,
      "accessing model": 836,
      "furthermore present": 12008,
      "internal representation": 15440,
      "representation particular": 28148,
      "perform multiple": 24495,
      "answering mcqa": 2064,
      "datasets english": 7103,
      "literature recent": 18044,
      "high school": 13582,
      "studies mainly": 31275,
      "mainly focused": 19404,
      "aim create": 1639,
      "novel highquality": 23086,
      "physics chemistry": 24943,
      "used evaluate": 34598,
      "llms smaller": 18955,
      "context question": 5911,
      "wellknown llms": 35526,
      "proposed dataset": 26596,
      "shows promising": 29934,
      "research purposes": 28353,
      "models unlock": 22096,
      "crucial human": 6440,
      "human reasoning": 13860,
      "firstofitskind largescale": 11566,
      "pairs diverse": 23977,
      "set tests": 29709,
      "sentence embedding": 29532,
      "embedding models": 9131,
      "chatgpt llama": 4486,
      "llama chatgpt": 18084,
      "accuracy humans": 888,
      "humans furthermore": 13924,
      "furthermore observe": 12007,
      "performance zeroshot": 24818,
      "model transparency": 20843,
      "societal impact": 30438,
      "models growing": 21343,
      "help improve": 13506,
      "time introduce": 33129,
      "models spanning": 21995,
      "build foundation": 3910,
      "details model": 7846,
      "downstream use": 8702,
      "model developers": 20467,
      "model gpt4": 20559,
      "gpt4 openai": 13101,
      "llama meta": 18124,
      "number users": 23173,
      "stateoftheart large": 30939,
      "testing llms": 32806,
      "llms diverse": 18544,
      "diverse sizes": 8463,
      "finetuning evaluation": 11399,
      "deployment llms": 7655,
      "using recent": 34899,
      "stateoftheart techniques": 30997,
      "tuning using": 34019,
      "llama using": 18149,
      "align model": 1734,
      "behaviors human": 3327,
      "instructiontuned model": 15298,
      "responses paper": 28503,
      "instructiontuned llm": 15295,
      "ranking approaches": 27230,
      "teacher llm": 32582,
      "allows model": 1814,
      "using contextual": 34759,
      "contextual understanding": 5955,
      "stronger llms": 31198,
      "llms furthermore": 18622,
      "furthermore apply": 11987,
      "super natural": 31637,
      "natural instructions": 22506,
      "test tasks": 32791,
      "obtain better": 23247,
      "baselines code": 3263,
      "teaching language": 32591,
      "models selfimprove": 21960,
      "significant recent": 30018,
      "learn smaller": 17513,
      "ability approach": 656,
      "approach improve": 2296,
      "performance math": 24672,
      "contrast prior": 6015,
      "collect feedback": 4916,
      "model experiments": 20501,
      "mt5 shown": 22257,
      "face difficulties": 10823,
      "particularly lowresource": 24350,
      "effective crosslingual": 8863,
      "address limitations": 1343,
      "limitations present": 17930,
      "present new": 25541,
      "new powerful": 22832,
      "universal dependencies": 34371,
      "able learn": 747,
      "syntactic tree": 31827,
      "method consistently": 19892,
      "different language": 8087,
      "performance advantage": 24520,
      "holds true": 13742,
      "distribution shifts": 8397,
      "test datasets": 32765,
      "generalization capacity": 12210,
      "model especially": 20493,
      "especially important": 9738,
      "wide adoption": 35545,
      "adoption pretrained": 1413,
      "pretrained foundational": 25646,
      "remains poorly": 28008,
      "learning tl": 17664,
      "present case": 25517,
      "pretrained foundation": 25644,
      "models encode": 21198,
      "different representations": 8132,
      "labeled training": 15958,
      "model interpretation": 20592,
      "generating instructiontuning": 12432,
      "instructiontuning data": 15302,
      "using incontext": 34796,
      "generation techniques": 12619,
      "techniques selfinstruct": 32663,
      "wang et": 35415,
      "limitation approaches": 17910,
      "models 175b": 20924,
      "application techniques": 2138,
      "models smaller": 21989,
      "permissive licenses": 24864,
      "approach effective": 2265,
      "lm outputs": 19058,
      "select highquality": 29378,
      "highquality synthetic": 13700,
      "algorithm leverages": 1709,
      "empirical investigations": 9229,
      "method yields": 19986,
      "higherquality instruction": 13617,
      "improves performances": 14390,
      "significant margins": 29995,
      "lms generate": 19086,
      "generate useful": 12337,
      "models biased": 21027,
      "gpt4 various": 13127,
      "prompting technique": 26397,
      "used llms": 34609,
      "furthermore work": 12016,
      "insights potential": 15081,
      "biases induced": 3674,
      "spatial understanding": 30643,
      "understanding large": 34239,
      "llms remarkable": 18908,
      "capabilities variety": 4079,
      "text training": 32957,
      "navigation tasks": 22597,
      "llms particular": 18844,
      "represent reason": 28134,
      "reason spatial": 27358,
      "aspects spatial": 2580,
      "tuned large": 33960,
      "demonstrate remarkable": 7488,
      "despite numerous": 7797,
      "performance instructiontuned": 24633,
      "remains lack": 27998,
      "lack comprehensive": 15979,
      "comprehensive investigation": 5382,
      "gap present": 12102,
      "benchmark specifically": 3411,
      "covering 13": 6323,
      "primary categories": 25917,
      "language detection": 16062,
      "emotion recognition": 9202,
      "12 language": 91,
      "writing scripts": 35856,
      "various multilingual": 35122,
      "finetuning zeroshot": 11557,
      "learning comprehensive": 17554,
      "comprehensive analysis": 5353,
      "reveals existing": 28815,
      "opensource instruction": 23504,
      "tuned llms": 33961,
      "llms struggle": 18972,
      "struggle understand": 31246,
      "close random": 4673,
      "baseline cases": 3243,
      "chatgpt outperforms": 4492,
      "benchmark available": 3356,
      "content using": 5875,
      "data paper": 6791,
      "powerful pretrained": 25351,
      "information target": 14916,
      "content information": 5863,
      "information input": 14873,
      "sentence respectively": 29542,
      "respectively compared": 28456,
      "embeddings used": 9148,
      "used previous": 34618,
      "richer information": 28877,
      "furthermore adopt": 11985,
      "provides effective": 26750,
      "model construct": 20441,
      "informative prefixes": 14933,
      "performance evaluations": 24586,
      "subjective evaluations": 31425,
      "benchmark evaluating": 3385,
      "modeling evaluation": 20892,
      "current landscape": 6496,
      "llama mistral": 18126,
      "benchmarks focus": 3442,
      "fundamental linguistic": 11979,
      "paper advocate": 24007,
      "valuable tool": 35020,
      "findings shed": 11253,
      "word representations": 35647,
      "representations learning": 28167,
      "learning mechanisms": 17600,
      "new opportunities": 22827,
      "approach provide": 2329,
      "pretraining complex": 25788,
      "understand reason": 34201,
      "require reasoning": 28224,
      "temporal dependencies": 32695,
      "make use": 19487,
      "graph structure": 13230,
      "relations sentences": 27870,
      "t5 multiple": 31958,
      "various settings": 35143,
      "settings analysis": 29730,
      "code pretrained": 4796,
      "data learning": 6755,
      "open information": 23395,
      "involves extracting": 15629,
      "techniques offer": 32653,
      "offer unique": 23294,
      "original sentence": 23720,
      "require significant": 28225,
      "task form": 32127,
      "model reducing": 20747,
      "extensive training": 10715,
      "data furthermore": 6719,
      "introduce innovative": 15510,
      "reducing training": 27759,
      "time experimental": 33123,
      "previous sota": 25878,
      "sota methods": 30534,
      "comparable results": 5089,
      "design automated": 7697,
      "task automatic": 32082,
      "realworld dataset": 27337,
      "discourse structure": 8265,
      "extensive automatic": 10679,
      "automatic human": 2884,
      "human experiments": 13814,
      "framework outperforms": 11886,
      "target audience": 32046,
      "producing coherent": 26165,
      "improves large": 14380,
      "evaluation generation": 9956,
      "llms frequently": 18621,
      "multiple aspects": 22378,
      "lack coherence": 15978,
      "challenging natural": 4388,
      "base llm": 3121,
      "task multiple": 32162,
      "tasks llm": 32407,
      "llm response": 18362,
      "generation evaluate": 12493,
      "correctness consistency": 6207,
      "outperform gpt4": 23775,
      "constraint satisfaction": 5786,
      "retrieval augmented": 28733,
      "augmented large": 2819,
      "cost reduction": 6253,
      "commercial large": 4984,
      "ability generating": 677,
      "retrieval augmentation": 28732,
      "retrievalaugmented llms": 28764,
      "input token": 15034,
      "size llms": 30261,
      "llms mitigate": 18801,
      "mitigate propose": 20258,
      "finetuned datasets": 11306,
      "datasets generated": 7121,
      "varying lengths": 35175,
      "size removing": 30280,
      "adequately evaluate": 1383,
      "effectiveness proposed": 8964,
      "methods propose": 20080,
      "propose utilize": 26585,
      "dataset called": 6945,
      "improvement accuracy": 14325,
      "provides flexible": 26753,
      "size performance": 30270,
      "factual recall": 10889,
      "memorized pretraining": 19798,
      "pretraining new": 25826,
      "new information": 22808,
      "using counterfactual": 34761,
      "data method": 6766,
      "method increase": 19934,
      "increase rate": 14604,
      "rate generating": 27266,
      "simply scaling": 30183,
      "work contributes": 35683,
      "body evidence": 3804,
      "proof concept": 26460,
      "control model": 6054,
      "excellent generalization": 10161,
      "abilities pretrained": 636,
      "direct training": 8218,
      "data making": 6765,
      "making better": 19498,
      "better foundation": 3604,
      "transfer knowledge": 33674,
      "source domain": 30557,
      "domain target": 8596,
      "fail account": 10899,
      "root cause": 28970,
      "source data": 30556,
      "data distribution": 6679,
      "domains study": 8641,
      "study proposes": 31383,
      "plms finetuning": 25046,
      "model feature": 20517,
      "feature extractor": 11024,
      "adversarial loss": 1516,
      "loss designed": 19243,
      "designed improve": 7734,
      "representation learning": 28141,
      "training compared": 33454,
      "extracted features": 10749,
      "processing computer": 26099,
      "vision downstream": 35293,
      "existing instructiontuning": 10279,
      "instructiontuning datasets": 15304,
      "datasets suffer": 7176,
      "helpful responses": 13517,
      "specific fields": 30694,
      "question ensure": 27067,
      "ensure comprehensive": 9602,
      "comprehensive coverage": 5362,
      "balanced distribution": 3084,
      "set covering": 29680,
      "realworld questions": 27343,
      "finetune llama": 11288,
      "gpt4 human": 13084,
      "win rate": 35610,
      "performance generative": 24617,
      "generative modeling": 12675,
      "tasks diffusion": 32298,
      "diffusion models": 8184,
      "models rely": 21924,
      "work bridge": 35672,
      "gap proposing": 12107,
      "integrates seamlessly": 15329,
      "significantly boosts": 30040,
      "boosts performance": 3829,
      "test score": 32783,
      "language diffusion": 16064,
      "generates faithful": 12400,
      "techniques like": 32647,
      "temperature scaling": 32687,
      "enables controllable": 9295,
      "left right": 17692,
      "pretraining architectures": 25785,
      "encompass various": 9386,
      "including autoencoding": 14456,
      "autoencoding models": 2847,
      "model long": 20630,
      "vector quantization": 35195,
      "seamlessly integrates": 29305,
      "point cloud": 25063,
      "model versatile": 20860,
      "tasks leading": 32395,
      "leading improved": 17475,
      "highly competitive": 13660,
      "results unconditional": 28700,
      "achieves superior": 1076,
      "different downstream": 8073,
      "tasks boost": 32255,
      "lms generation": 19088,
      "coherent contextually": 4891,
      "contextually relevant": 5970,
      "generated outputs": 12377,
      "outputs lack": 23892,
      "finetuning entire": 11396,
      "specifically construct": 30732,
      "benchmarks human": 3447,
      "consistently leads": 5753,
      "outputs language": 23893,
      "general mechanism": 12178,
      "sufficiently large": 31564,
      "llama families": 18099,
      "using causal": 34748,
      "causal interventions": 4239,
      "id vectors": 13982,
      "vectors corresponding": 35198,
      "symbolic knowledge": 31801,
      "knowledge incontext": 15864,
      "providing step": 26783,
      "step understanding": 31050,
      "understanding general": 34223,
      "generalpurpose ai": 12246,
      "llm evaluations": 18298,
      "ability ai": 655,
      "ai agent": 1594,
      "plays important": 25028,
      "role human": 28957,
      "2023 work": 237,
      "asks llm": 2561,
      "llm produce": 18352,
      "produce text": 26155,
      "skills number": 30314,
      "high probability": 13578,
      "text significantly": 32939,
      "different text": 8150,
      "paper develops": 24035,
      "llama2 70b": 18160,
      "70b model": 505,
      "version popular": 35233,
      "popular llm": 25122,
      "llm leaderboards": 18332,
      "reasonable performance": 27360,
      "ecosystem open": 8816,
      "future models": 12039,
      "challenging existing": 4384,
      "benchmarks metrics": 3462,
      "address problem": 1349,
      "llms efficiently": 18554,
      "benchmarks propose": 3468,
      "propose comprehensive": 26500,
      "scales 7b": 29151,
      "7b 13b": 525,
      "13b 33b": 120,
      "parameters conduct": 24233,
      "conduct systematic": 5623,
      "capabilities behaviors": 4003,
      "analyze key": 1996,
      "finetuning llm": 11447,
      "position bias": 25181,
      "techniques including": 32645,
      "obtains stateoftheart": 23263,
      "proposed new": 26615,
      "a100 gpus": 618,
      "high agreement": 13550,
      "answer multimodal": 2047,
      "multiturn chat": 22461,
      "targeted data": 32064,
      "models rapid": 21892,
      "rapid advancement": 27239,
      "advancement large": 1454,
      "data synthesis": 6886,
      "synthetic datasets": 31855,
      "datasets synthetic": 7178,
      "suffer lack": 31551,
      "lack diversity": 15983,
      "datasets utilizing": 7188,
      "various language": 35105,
      "encoderonly encoderdecoder": 9376,
      "encoderdecoder decoderonly": 9365,
      "decoderonly models": 7263,
      "original training": 23723,
      "data vs": 6910,
      "dataset compared": 6952,
      "dataset demonstrates": 6971,
      "complexity diversity": 5301,
      "yields impressive": 35923,
      "human efforts": 13804,
      "benchmarks large": 3452,
      "representative models": 28185,
      "including open": 14509,
      "specifically consider": 30731,
      "feasibility using": 11017,
      "llms estimate": 18569,
      "public opinion": 26841,
      "introduction new": 15555,
      "survey existing": 31773,
      "practical implementation": 25366,
      "potential risks": 25296,
      "retrieval knowledge": 28743,
      "graph construction": 13219,
      "entities texts": 9641,
      "face challenges": 10822,
      "challenges like": 4357,
      "manual effort": 19557,
      "poor generalization": 25103,
      "pressing need": 25617,
      "efficiently extract": 9071,
      "study explores": 31331,
      "potential pretrained": 25289,
      "articles using": 2525,
      "prompts propose": 26435,
      "strategy called": 31116,
      "prompts based": 26405,
      "construct datasets": 5796,
      "prompts test": 26444,
      "llms experimental": 18586,
      "llms significant": 18947,
      "furthermore analyze": 11986,
      "factors influence": 10872,
      "influence performance": 14839,
      "llms know": 18733,
      "method large": 19937,
      "reveals llms": 28816,
      "method detect": 19900,
      "questions llm": 27121,
      "llm does": 18292,
      "prone generate": 26453,
      "corresponding answers": 6225,
      "generated answers": 12342,
      "questions model": 27124,
      "released llms": 27925,
      "prompts tasks": 26443,
      "tasks remains": 32480,
      "task context": 32100,
      "context representation": 5913,
      "improving llms": 14413,
      "evaluate efficacy": 9834,
      "efficacy approach": 8986,
      "finetuning based": 11376,
      "opensource llama2": 23516,
      "models web": 22125,
      "tasks additionally": 32235,
      "performance prior": 24721,
      "stepbystep instructions": 31054,
      "web agents": 35472,
      "despite immense": 7786,
      "immense promise": 14116,
      "promise performing": 26277,
      "tasks theoretical": 32531,
      "understanding limitations": 34245,
      "study generalization": 31334,
      "unseen datasets": 34435,
      "datasets recent": 7167,
      "work shown": 35779,
      "randomly initialized": 27181,
      "closed form": 4676,
      "form results": 11741,
      "experimental evidence": 10386,
      "evidence corroborates": 10060,
      "makes approach": 19489,
      "motivating need": 22232,
      "current work": 6539,
      "logistic regression": 19160,
      "certain cases": 4272,
      "shown superior": 29918,
      "theoretical grounding": 33050,
      "framework suggests": 11896,
      "performance classification": 24543,
      "using conventional": 34760,
      "used improve": 34605,
      "regression tasks": 27823,
      "evaluation question": 9998,
      "numerous applications": 23182,
      "model aid": 20360,
      "model prompt": 20731,
      "research finetuned": 28316,
      "finetuned pretrained": 11351,
      "distilbert model": 8331,
      "training transformer": 33640,
      "applied generate": 2189,
      "using llama": 34820,
      "questions squad": 27133,
      "squad dataset": 30842,
      "effectiveness different": 8942,
      "achieved high": 1008,
      "exhibit impressive": 10221,
      "impressive reasoning": 14246,
      "reasoning data": 27400,
      "tasks small": 32508,
      "surpassing models": 31757,
      "models 100b": 20918,
      "augmentation ability": 2796,
      "nlp datasets": 22929,
      "student models": 31256,
      "brought significant": 3895,
      "structure transformer": 31216,
      "lack explicit": 15989,
      "syntactic generalization": 31816,
      "selfattention layer": 29408,
      "syntactic language": 31819,
      "new tokens": 22859,
      "replacement standard": 28102,
      "tasks building": 32256,
      "summarization systems": 31624,
      "practical perspective": 25368,
      "perspective paper": 24897,
      "paper studies": 24136,
      "effectively build": 8913,
      "realworld usage": 27350,
      "gpt 35": 12842,
      "closedsource llms": 4687,
      "generally better": 12241,
      "terms performance": 32748,
      "smaller opensource": 30392,
      "llama 7b": 18066,
      "13b achieve": 123,
      "comparable large": 5080,
      "associated using": 2653,
      "finetuned versions": 11363,
      "achieve competitive": 954,
      "balancing performance": 3087,
      "llama27b model": 18209,
      "paper offers": 24082,
      "practical insights": 25367,
      "insights using": 15084,
      "processing task": 26125,
      "tackle task": 32002,
      "llms solve": 18959,
      "using diverse": 34769,
      "range llms": 27197,
      "gpt4 opt": 13105,
      "experiments encompass": 10439,
      "learning settings": 17654,
      "settings evaluate": 29732,
      "critical challenge": 6383,
      "human value": 13874,
      "frequently overlooked": 11928,
      "online platforms": 23367,
      "leverage user": 17762,
      "online sources": 23370,
      "google play": 12829,
      "end users": 9422,
      "users using": 34704,
      "theory approach": 33057,
      "major categories": 19438,
      "human values": 13875,
      "stateoftheart pretrained": 30977,
      "depending data": 7620,
      "performed best": 24827,
      "recommendations used": 27653,
      "software developers": 30453,
      "evolving needs": 10082,
      "metrics measuring": 20143,
      "optimize quantization": 23643,
      "quantization large": 27004,
      "raised concerns": 27165,
      "deployment need": 7657,
      "need llm": 22636,
      "study introduces": 31344,
      "compressed llms": 5405,
      "deeper insights": 7348,
      "llama2 model": 18186,
      "model family": 20515,
      "evaluations indicate": 10031,
      "analysis method": 1940,
      "aimed identifying": 1653,
      "related specific": 27856,
      "specific aspects": 30683,
      "framework successfully": 11895,
      "learning prompting": 17636,
      "prompting approach": 26372,
      "approach enhance": 2270,
      "enhance model": 9519,
      "application approach": 2126,
      "research aims": 28289,
      "models study": 22019,
      "model developed": 20466,
      "model utilizing": 20857,
      "extraction task": 10773,
      "aspect opinion": 2567,
      "aspect term": 2568,
      "model specifically": 20805,
      "learning train": 17665,
      "steering llms": 31029,
      "llms similar": 18952,
      "userspecified information": 34706,
      "method allows": 19876,
      "small subset": 30369,
      "model attention": 20379,
      "changing model": 4415,
      "user instructions": 34655,
      "instructions integrate": 15255,
      "new knowledge": 22810,
      "user inputs": 34654,
      "inputs leading": 15049,
      "improvement variety": 14351,
      "tasks average": 32249,
      "improvement 22": 14324,
      "llama7b code": 18232,
      "finetuning code": 11381,
      "coding capabilities": 4866,
      "finetuning approaches": 11373,
      "approaches typically": 2399,
      "tailored specific": 32018,
      "specific downstream": 30691,
      "finetuning task": 11543,
      "requiring extensive": 28270,
      "posing challenges": 25178,
      "challenges terms": 4377,
      "deployment maintenance": 7656,
      "coderelated tasks": 4847,
      "overcome limitations": 23923,
      "tasks incorporating": 32372,
      "varying difficulty": 35172,
      "difficulty levels": 8177,
      "convergence speeds": 6086,
      "finetuning single": 11526,
      "efficient training": 9061,
      "capabilities including": 4025,
      "resulting significantly": 28561,
      "achieves impressive": 1049,
      "benchmark surpassing": 3415,
      "surpassing gpt4": 31755,
      "gpt4 performance": 13111,
      "make recommendations": 19481,
      "user satisfaction": 34671,
      "task detecting": 32110,
      "facilitate development": 10837,
      "manually annotate": 19562,
      "use evaluate": 34526,
      "rapidly growing": 27257,
      "innovative solution": 14998,
      "leverages opensource": 17771,
      "llms llama2": 18774,
      "ensure data": 9603,
      "augmented generation": 2815,
      "generation rag": 12587,
      "optimization dpo": 23624,
      "extensive experimentation": 10692,
      "comprising 10000": 5434,
      "qa pairs": 26913,
      "pairs preference": 23982,
      "preference data": 25467,
      "data demonstrate": 6671,
      "improvement quality": 14345,
      "quality answers": 26940,
      "contributions include": 6040,
      "development novel": 7965,
      "human assessments": 13795,
      "challenges future": 4346,
      "qa platform": 26914,
      "free lunch": 11910,
      "models retraining": 21942,
      "pretrained parameters": 25741,
      "versatile plugandplay": 35228,
      "models mitigating": 21730,
      "model parameter": 20687,
      "encoder decoderbased": 9347,
      "parameter value": 24204,
      "multiple taskspecific": 22424,
      "zeroshot accuracy": 35951,
      "performance merged": 24674,
      "llms proposed": 18877,
      "proposed recent": 26618,
      "opensource ones": 23538,
      "new records": 22838,
      "development llms": 7961,
      "continual pretraining": 5979,
      "forgetting issues": 11731,
      "llms important": 18684,
      "enlarging model": 9583,
      "model 13": 20332,
      "llama2 foundation": 18177,
      "different stages": 8142,
      "work language": 35730,
      "models little": 21463,
      "little understanding": 18053,
      "new models": 22825,
      "reflect differences": 27789,
      "differences model": 8037,
      "wild work": 35608,
      "models revealing": 21945,
      "revealing shared": 28811,
      "input perturbations": 15019,
      "designed target": 7745,
      "specific linguistic": 30702,
      "framework offers": 11885,
      "increase size": 14606,
      "size pretraining": 30279,
      "relatively better": 27885,
      "better understood": 3635,
      "gpt2 experiments": 12890,
      "experiments observe": 10464,
      "observe large": 23230,
      "encoded large": 9340,
      "key reason": 15784,
      "samples large": 29080,
      "increasingly trained": 14645,
      "finetuning datasets": 11386,
      "datasets data": 7089,
      "benchmark data": 3370,
      "data methods": 6767,
      "test data": 32763,
      "benchmark achieve": 3351,
      "used benchmarks": 34587,
      "used pretraining": 34617,
      "revealing significant": 28812,
      "previously unknown": 25903,
      "humaneval benchmark": 13897,
      "suggesting potential": 31584,
      "potential risk": 25295,
      "urge community": 34493,
      "approaches using": 2402,
      "community actively": 5057,
      "generation approach": 12458,
      "labelled data": 15964,
      "dataset extracted": 6989,
      "diversity metrics": 8480,
      "balance diversity": 3078,
      "diversity selected": 8483,
      "manual curation": 19556,
      "expected output": 10352,
      "different strategies": 8143,
      "evaluated performance": 9881,
      "open large": 23400,
      "addition evaluation": 1239,
      "evaluated models": 9880,
      "exhibited substantial": 10239,
      "noisy data": 22987,
      "data provide": 6816,
      "provide best": 26686,
      "model endtoend": 20488,
      "generated synthetic": 12390,
      "humanauthored text": 13888,
      "applicability large": 2120,
      "text remains": 32931,
      "remains unexplored": 28021,
      "unexplored study": 34303,
      "study addresses": 31295,
      "models conduct": 21088,
      "pubmed abstracts": 26874,
      "using different": 34767,
      "parameter sizes": 24199,
      "size grows": 30251,
      "outputs future": 23889,
      "enhancing model": 9568,
      "boosting large": 3820,
      "t0 flan": 31932,
      "exhibit remarkable": 10226,
      "remarkable generalization": 28042,
      "abilities unseen": 644,
      "sizes ranging": 30303,
      "ranging billion": 27219,
      "substantial computational": 31460,
      "resources making": 28439,
      "models downstream": 21172,
      "applications particularly": 2171,
      "approaches prompt": 2386,
      "potential address": 25236,
      "challenges introduce": 4352,
      "tasks serve": 32497,
      "boosting performance": 3825,
      "11 language": 72,
      "magnitude larger": 19383,
      "margin furthermore": 19585,
      "including finetuning": 14477,
      "finetuning incontext": 11418,
      "performance enhancement": 24581,
      "demonstrated superior": 7551,
      "understanding abilities": 34206,
      "abilities including": 626,
      "including zeroshot": 14530,
      "capabilities similar": 4069,
      "human ones": 13848,
      "ones study": 23349,
      "text conducted": 32834,
      "selfpaced reading": 29428,
      "significantly longer": 30067,
      "reading times": 27306,
      "humans easily": 13922,
      "tested variety": 32798,
      "extent models": 10723,
      "recent llms": 27531,
      "gpt3 vicuna": 13010,
      "correlate human": 6212,
      "fail predict": 10904,
      "knowledge code": 15824,
      "languages modalities": 16897,
      "resulting significant": 28560,
      "progress natural": 26218,
      "tasks consequently": 32275,
      "evaluation research": 9999,
      "english language": 9482,
      "relatively unexplored": 27896,
      "new llms": 22819,
      "languages study": 16915,
      "aims expand": 1663,
      "including new": 14506,
      "new datasets": 22791,
      "languages including": 16878,
      "evaluate stateoftheart": 9865,
      "gpt4 palm2": 13110,
      "additionally include": 1289,
      "multimodal datasets": 22341,
      "datasets benchmark": 7069,
      "benchmark assess": 3355,
      "tasks notably": 32432,
      "vice versa": 35244,
      "data contamination": 6659,
      "obtain accurate": 23246,
      "assessment llm": 2618,
      "languages using": 16920,
      "improve robustness": 14295,
      "robustness incontext": 28947,
      "inference recent": 14804,
      "studies demonstrated": 31265,
      "demonstrated large": 7532,
      "taskspecific prompts": 32568,
      "examples existing": 10123,
      "existing literature": 10288,
      "adversarial inputs": 1515,
      "enhanced performance": 9535,
      "performance observed": 24699,
      "explanations nles": 10542,
      "robustness llms": 28948,
      "datasets introduce": 7134,
      "prompting llm": 26386,
      "llm chatgpt": 18281,
      "evaluate popular": 9859,
      "llms gpt35turbo": 18653,
      "llama2 vicuna": 18196,
      "yields improvement": 35924,
      "improvement icl": 14341,
      "prompt selection": 26341,
      "shown significantly": 29916,
      "improve icl": 14269,
      "llms given": 18641,
      "labeled examples": 15956,
      "examples input": 10129,
      "llm learns": 18333,
      "learns perform": 17679,
      "models supervised": 22031,
      "represent underlying": 28136,
      "icl using": 13979,
      "using test": 34927,
      "test case": 32761,
      "understanding experiments": 34222,
      "based simple": 3223,
      "syntactic transformations": 31826,
      "requires accurate": 28246,
      "improved chainofthought": 14307,
      "model provided": 20735,
      "intermediate computation": 15425,
      "performed experiments": 24829,
      "gpt palm": 12861,
      "linguistic phenomenon": 18018,
      "pretraining corpus": 25790,
      "pretrained code": 25637,
      "greater extent": 13269,
      "languages recent": 16910,
      "proliferation large": 26258,
      "widely adopted": 35569,
      "present analysis": 25511,
      "analysis popular": 1947,
      "popular large": 25119,
      "llama gpt4": 18111,
      "news topic": 22889,
      "topic classification": 33282,
      "highresource languages": 13707,
      "like english": 17860,
      "tasks gpt4": 32351,
      "gpt4 average": 13057,
      "generative tasks": 12706,
      "stateoftheart supervised": 30994,
      "findings present": 11244,
      "languages represented": 16913,
      "growing popularity": 13317,
      "popularity large": 25142,
      "model temporal": 20826,
      "tasks reflecting": 32477,
      "gap limited": 12095,
      "contrary expectations": 6008,
      "does guarantee": 8527,
      "temporal information": 32697,
      "information sentence": 14912,
      "available pretraining": 2997,
      "tasks conclude": 32271,
      "llms lack": 18740,
      "llm outputs": 18343,
      "advances transformerbased": 1488,
      "transformerbased large": 33751,
      "models great": 21340,
      "great strides": 13259,
      "strides natural": 31148,
      "models nonautoregressive": 21756,
      "benchmarks work": 3477,
      "study scientific": 31394,
      "domains large": 8623,
      "incontext demonstrations": 14543,
      "learn new": 17511,
      "groundtruth labels": 13294,
      "labels address": 15967,
      "llms follow": 18618,
      "present different": 25525,
      "types factual": 34061,
      "prompts zeroshot": 26448,
      "sentence classification": 29528,
      "performance larger": 24650,
      "70b parameters": 506,
      "parameters limited": 24268,
      "alignment methods": 1772,
      "methods finally": 20038,
      "evaluation method": 9971,
      "models leading": 21438,
      "leading proprietary": 17483,
      "augmenting language": 2824,
      "model lm": 20629,
      "underlying reasons": 34162,
      "elusive work": 9117,
      "mlp layer": 20304,
      "advanced model": 1431,
      "information training": 14920,
      "gpt2 117m": 12867,
      "performance setting": 24749,
      "datasets provide": 7163,
      "support downstream": 31706,
      "grounded given": 13288,
      "knowledge introduce": 15867,
      "knowledge grounded": 15861,
      "model t5large": 20822,
      "novel challenges": 23066,
      "human experts": 13817,
      "educational domain": 8841,
      "samples prompt": 29086,
      "prompt inputs": 26330,
      "inputs exploring": 15048,
      "effective incontext": 8874,
      "sampling llm": 29092,
      "llm fewshot": 18304,
      "fewshot prompt": 11118,
      "set data": 29681,
      "data samples": 6845,
      "inside single": 15063,
      "single prompt": 30220,
      "design leverage": 7709,
      "multiple prompt": 22412,
      "improve llm": 14273,
      "prediction results": 25434,
      "icl prompt": 13978,
      "sota llms": 30533,
      "consistently enhance": 5748,
      "study suggests": 31400,
      "strategy improve": 31123,
      "sheds light": 29800,
      "light new": 17830,
      "new promising": 22834,
      "promising future": 26288,
      "political bias": 25094,
      "llms extensive": 18596,
      "llms learned": 18750,
      "offensive toxic": 23284,
      "language recent": 16816,
      "gpt language": 12849,
      "models recognize": 21912,
      "generated content": 12348,
      "algorithm allows": 1701,
      "generating harmful": 12426,
      "syntactic properties": 31822,
      "aim contribute": 1638,
      "ongoing effort": 23357,
      "humanai interaction": 13883,
      "multitask model": 22452,
      "handle multiple": 13410,
      "training commonly": 33452,
      "diverse contexts": 8417,
      "long sequences": 19178,
      "length usually": 17713,
      "computation efficient": 5447,
      "enable efficient": 9286,
      "using dynamic": 34770,
      "approach handle": 2292,
      "execution time": 10201,
      "enabling highly": 9322,
      "training extensive": 33519,
      "training t5": 33627,
      "training gpt": 33524,
      "augmented language": 2817,
      "models increasing": 21389,
      "models computation": 21081,
      "learning capacity": 17549,
      "style models": 31413,
      "approach dubbed": 2264,
      "large set": 17276,
      "play role": 25019,
      "t5 family": 31944,
      "tasks similar": 32505,
      "approaches require": 2391,
      "languages llms": 16892,
      "processes llms": 26092,
      "train new": 33370,
      "datasets metrics": 7147,
      "metrics used": 20150,
      "settings paper": 29741,
      "aforementioned challenges": 1552,
      "introduce multilingual": 15516,
      "languages propose": 16908,
      "llms new": 18817,
      "model performed": 20708,
      "highresource language": 13706,
      "performance instruction": 24632,
      "promising method": 26289,
      "method creating": 19896,
      "multilingual llms": 22318,
      "llms lowresource": 18787,
      "model adapters": 20356,
      "lora adapters": 19228,
      "task generalization": 32129,
      "generalization paper": 12225,
      "models arbitrary": 20987,
      "increasing compute": 14620,
      "requirements training": 28243,
      "inference results": 14806,
      "outperforms base": 23805,
      "individual models": 14717,
      "finetuned tasks": 11357,
      "tasks best": 32253,
      "inference code": 14764,
      "code study": 4818,
      "study available": 31304,
      "word representation": 35646,
      "representation integrates": 28140,
      "pretrained word": 25781,
      "drawing recent": 8728,
      "construct novel": 5804,
      "eliminates need": 9108,
      "need backpropagation": 22621,
      "leveraging contextual": 17780,
      "dimensionality reduction": 8198,
      "techniques based": 32629,
      "based unigram": 3236,
      "providing strong": 26784,
      "strong interpretability": 31176,
      "twostep process": 34050,
      "critically relies": 6400,
      "contextually rich": 5971,
      "representations word": 28179,
      "partofspeech pos": 24366,
      "assess competitiveness": 2590,
      "explore applicability": 10572,
      "lm training": 19063,
      "training finetuning": 33522,
      "experiments illustrate": 10448,
      "conventional training": 6080,
      "t5 opt": 31959,
      "enhancement transfer": 9542,
      "teaching small": 32594,
      "outperform conventional": 23770,
      "conventional instructiontuned": 6074,
      "models benchmarks": 21013,
      "improved training": 14319,
      "training signals": 33615,
      "capable models": 4113,
      "teach small": 32580,
      "strategies different": 31103,
      "potentially different": 25311,
      "provide direct": 26695,
      "direct answer": 8210,
      "teach model": 32579,
      "model various": 20859,
      "various reasoning": 35139,
      "aim help": 1644,
      "model learn": 20607,
      "task evaluate": 32114,
      "15 diverse": 141,
      "100 tasks": 41,
      "unique prompts": 34363,
      "significantly surpasses": 30088,
      "performance levels": 24655,
      "better models": 3612,
      "advanced reasoning": 1441,
      "abilities zeroshot": 649,
      "weights publicly": 35513,
      "support research": 31711,
      "evaluation alignment": 9917,
      "demonstrated exceptional": 7520,
      "designed specifically": 7744,
      "continual training": 5980,
      "model derived": 20460,
      "extensive data": 10681,
      "academic benchmarks": 785,
      "demonstrate ability": 7433,
      "ability general": 670,
      "abstract generation": 768,
      "highperformance computing": 13675,
      "computing large": 5514,
      "model exhibited": 20498,
      "efficacy various": 8994,
      "computing hpc": 5513,
      "interpret model": 15455,
      "model responses": 20755,
      "responses response": 28510,
      "response challenge": 28472,
      "challenge propose": 4327,
      "using generated": 34784,
      "generated qa": 12379,
      "qa questionanswer": 26915,
      "questionanswer instances": 27079,
      "hpc tasks": 13771,
      "demonstrate comparable": 7439,
      "experiments opensource": 10465,
      "extensive results": 10712,
      "bridge performance": 3869,
      "gap llms": 12096,
      "pave way": 24417,
      "way llms": 35443,
      "utilization language": 34953,
      "computing applications": 5512,
      "models model": 21735,
      "model merging": 20638,
      "models continually": 21106,
      "result significant": 28547,
      "overcome problem": 23924,
      "enables finetuned": 9299,
      "method conducted": 19891,
      "form model": 11740,
      "finetuned language": 11321,
      "pretrained base": 25628,
      "models domains": 21169,
      "despite simplicity": 7814,
      "surprisingly effective": 31767,
      "able achieve": 741,
      "strong empirical": 31169,
      "empirical performance": 9231,
      "domain conduct": 8557,
      "experiments llama": 10458,
      "results validate": 28704,
      "method code": 19888,
      "code checkpoints": 4725,
      "checkpoints available": 4522,
      "model inversion": 20594,
      "prompt tokens": 26350,
      "nexttoken probabilities": 22896,
      "surprising information": 31765,
      "preceding text": 25387,
      "prompts given": 26421,
      "model access": 20340,
      "llama2 7b": 18162,
      "code reproducing": 4813,
      "aims improve": 1668,
      "using prefix": 34877,
      "lora finetuning": 19231,
      "methods create": 20015,
      "qa data": 26907,
      "data based": 6621,
      "based prompt": 3211,
      "words given": 35657,
      "quickly obtain": 27136,
      "field provide": 11152,
      "provide data": 26692,
      "support finetuning": 31707,
      "training method": 33563,
      "tasks compared": 32269,
      "compared lora": 5148,
      "improves bleu": 14371,
      "bleu rouge": 3766,
      "rouge metrics": 28980,
      "compared model": 5149,
      "demonstrates effectiveness": 7557,
      "knowledge qa": 15895,
      "tasks provides": 32466,
      "provides new": 26759,
      "llms enhanced": 18565,
      "chat corpus": 4442,
      "corpus generation": 6183,
      "evaluation study": 10014,
      "introduces novel": 15544,
      "generator llm": 12717,
      "new samples": 22840,
      "quality assessment": 26941,
      "metric proposed": 20125,
      "sentence transformer": 29545,
      "translated english": 33814,
      "selfchat data": 29412,
      "processing techniques": 26130,
      "quality evaluation": 26957,
      "demonstrates significantly": 7570,
      "language comprehension": 16052,
      "resultant model": 28550,
      "establishes new": 9776,
      "substantial advancement": 31459,
      "special emphasis": 30656,
      "underrepresented languages": 34171,
      "learning assessment": 17541,
      "science requires": 29244,
      "solving programming": 30517,
      "developed using": 7935,
      "ai capability": 1601,
      "accuracy dramatically": 870,
      "academic integrity": 786,
      "achieving desired": 1090,
      "existing detection": 10269,
      "detection methods": 7871,
      "based similarity": 3222,
      "similarity metrics": 30131,
      "ai detection": 1606,
      "mixed success": 20271,
      "ai code": 1604,
      "need adapting": 22619,
      "models built": 21036,
      "built chatglm": 3931,
      "various ai": 35073,
      "social relationships": 30430,
      "outperforms mainstream": 23832,
      "large langauge": 16957,
      "langauge models": 16030,
      "including gpt": 14481,
      "especially terms": 9749,
      "subset training": 31453,
      "data facilitate": 6708,
      "dialogue generation": 8015,
      "falcon series": 10942,
      "open language": 23397,
      "180b parameters": 186,
      "trained diverse": 33393,
      "developed models": 7931,
      "cost making": 6248,
      "report detailed": 28113,
      "deep dive": 7317,
      "models permissive": 21812,
      "development open": 7966,
      "human cognition": 13800,
      "studies typically": 31289,
      "covers broad": 6331,
      "broad spectrum": 3887,
      "provides thorough": 26764,
      "experiments popular": 10466,
      "llama2 mistral": 18184,
      "indicate significant": 14695,
      "reasoning llms": 27420,
      "aviation domain": 3035,
      "llms demonstrating": 18531,
      "demonstrating exceptional": 7581,
      "domain resulting": 8591,
      "resulting low": 28555,
      "llms presents": 18863,
      "domain address": 8554,
      "built opensource": 3935,
      "continuously trained": 6002,
      "datasets experimental": 7110,
      "offers users": 23316,
      "users multiple": 34695,
      "provides accurate": 26747,
      "accurate contextually": 921,
      "research problems": 28348,
      "enhance efficiency": 9512,
      "llms crosslingual": 18499,
      "model input": 20584,
      "input layer": 15015,
      "language tokens": 16835,
      "different writing": 8160,
      "writing systems": 35860,
      "token represent": 33200,
      "effect pretraining": 8855,
      "multimodal language": 22349,
      "autonomous vehicles": 2932,
      "navigating complex": 22595,
      "complex realworld": 5288,
      "humanlike understanding": 13912,
      "novel visionlanguage": 23124,
      "visionlanguage model": 35313,
      "humanlike abilities": 13907,
      "multimodal inputs": 22346,
      "video image": 35265,
      "data text": 6892,
      "control signals": 6055,
      "provided instructions": 26740,
      "pretrained visionlanguage": 25777,
      "capabilities innovative": 4027,
      "dataset designed": 6973,
      "understanding intricate": 34236,
      "driving scenarios": 8738,
      "ability provide": 718,
      "comprehensive understanding": 5395,
      "understanding complex": 34216,
      "despite remarkable": 7809,
      "remarkable achievements": 28027,
      "linguistic bias": 18005,
      "regional languages": 27820,
      "series language": 29639,
      "specifically focuses": 30743,
      "sea languages": 29301,
      "built llama2": 3934,
      "model advanced": 20359,
      "continued pretraining": 5990,
      "better capture": 3596,
      "evaluation demonstrates": 9940,
      "large margins": 17230,
      "advancements generative": 1463,
      "showcasing effectiveness": 29848,
      "utilizing pretrained": 34976,
      "diverse applications": 8411,
      "applications despite": 2148,
      "despite potential": 7801,
      "capabilities lack": 4028,
      "lack adequate": 15975,
      "comprehensive benchmarks": 5358,
      "benchmarks particularly": 3467,
      "generalization performance": 12226,
      "performance multiple": 24683,
      "gaps introduce": 12117,
      "tailored evaluating": 32016,
      "study explore": 31330,
      "understanding instruction": 34234,
      "finetuning toxicity": 11549,
      "bias evaluation": 3644,
      "highquality humanannotated": 13689,
      "humanannotated datasets": 13885,
      "datasets instruction": 7133,
      "dataset covering": 6964,
      "cultures idioms": 6463,
      "classification question": 4606,
      "evaluations proposed": 10038,
      "proposed benchmark": 26595,
      "advancing language": 1493,
      "novel finetuning": 23080,
      "finetuning technique": 11545,
      "models involves": 21408,
      "method aims": 19875,
      "aims enhance": 1662,
      "enhance models": 9521,
      "demonstrating superior": 7590,
      "techniques yield": 32673,
      "improvement stateoftheart": 14347,
      "stateoftheart method": 30953,
      "models stronger": 22014,
      "stronger baseline": 31197,
      "baseline instruction": 3249,
      "current literature": 6509,
      "literature including": 18042,
      "research application": 28290,
      "strategies finetuning": 31106,
      "models approach": 20985,
      "significant step": 30026,
      "step direction": 31041,
      "improvement existing": 14338,
      "impact tokenization": 14139,
      "reason lies": 27357,
      "data hinders": 6727,
      "limiting potential": 17978,
      "investigate possibility": 15592,
      "addressing issue": 1374,
      "russian language": 29021,
      "language adaptation": 16035,
      "improves models": 14386,
      "models quality": 21880,
      "models demonstrates": 21138,
      "generate answers": 12263,
      "higher user": 13613,
      "user preference": 34663,
      "comprehensive survey": 5393,
      "survey large": 31774,
      "llms mainly": 18790,
      "textbased reasoning": 32975,
      "ability generalized": 673,
      "systematic review": 31874,
      "textattributed graphs": 32970,
      "utilizing llms": 34975,
      "encoder llm": 9351,
      "advantages disadvantages": 1504,
      "furthermore discuss": 11996,
      "applications methods": 2168,
      "datasets finally": 7116,
      "models additional": 20959,
      "training additional": 33438,
      "training explore": 33518,
      "llm key": 18326,
      "key findings": 15769,
      "effective knowledge": 8878,
      "knowledge integration": 15866,
      "models 7b": 20928,
      "13b 70b": 122,
      "knowledge study": 15912,
      "study highlights": 31338,
      "limitations incorporating": 17921,
      "information llms": 14884,
      "multimodal llms": 22361,
      "inference explicit": 14777,
      "llms empower": 18560,
      "multimodality understanding": 22370,
      "understanding capability": 34214,
      "capability semantic": 4103,
      "semantic generation": 29457,
      "reliance prompt": 27962,
      "generative nature": 12681,
      "tackle issue": 31997,
      "issue introduce": 15654,
      "inference method": 14792,
      "method prompt": 19959,
      "based highlighted": 3172,
      "autoregressive generation": 2939,
      "models highlighted": 21356,
      "attention weights": 2744,
      "weights leads": 35509,
      "llms vlms": 19042,
      "achieving impressive": 1095,
      "generation results": 12596,
      "training experiments": 33517,
      "experiments confirm": 10429,
      "confirm effectiveness": 5660,
      "input contexts": 15006,
      "llms recent": 18893,
      "recent surge": 27563,
      "falcon mistral": 10940,
      "provides diverse": 26749,
      "practitioners researchers": 25382,
      "code technical": 4821,
      "technical reports": 32612,
      "progress field": 26211,
      "training code": 33449,
      "intermediate results": 15432,
      "available community": 2969,
      "research making": 28336,
      "parameter llms": 24188,
      "pretrained scratch": 25745,
      "including training": 14522,
      "continually pushing": 5982,
      "llms opensource": 18832,
      "effort largescale": 9079,
      "stronger models": 31200,
      "addition introduce": 1241,
      "introduce contrastive": 15502,
      "innovative method": 14997,
      "responses inference": 28497,
      "token positions": 33197,
      "precise control": 25389,
      "gain deeper": 12056,
      "steers model": 31032,
      "highlevel concepts": 13623,
      "concepts represented": 5532,
      "era advanced": 9692,
      "experimental setup": 10409,
      "statistical model": 31019,
      "impact findings": 14122,
      "llms consistently": 18486,
      "forecasting models": 11723,
      "performance human": 24622,
      "human llm": 13842,
      "errors particularly": 9725,
      "careful consideration": 4169,
      "integrating llms": 15334,
      "llms practical": 18860,
      "llms useful": 19024,
      "best opensource": 3566,
      "having 50": 13473,
      "50 billion": 419,
      "llms comparing": 18477,
      "llm efficiently": 18294,
      "perform inference": 24493,
      "llama 70b": 18065,
      "base chat": 3114,
      "tools models": 33272,
      "models yield": 22139,
      "sota opensource": 30537,
      "models llama2": 21466,
      "leading performance": 17482,
      "performance major": 24670,
      "benchmarks leaderboards": 3453,
      "publicly releasing": 26865,
      "releasing models": 27931,
      "approach additional": 2231,
      "way making": 35444,
      "llms use": 19020,
      "annotations paper": 2032,
      "present innovative": 25537,
      "model called": 20406,
      "math problem": 19669,
      "achieved using": 1021,
      "automatically constructed": 2907,
      "manual annotation": 19554,
      "proximal policy": 26794,
      "policy optimization": 25086,
      "optimization ppo": 23633,
      "series opensource": 29643,
      "demonstrates exceptional": 7558,
      "accuracy enhanced": 874,
      "process supervision": 26085,
      "evolution llms": 10075,
      "chatgpt bard": 4458,
      "shown impressive": 29887,
      "meaningful responses": 19725,
      "propose lightweight": 26524,
      "specifically finetuned": 30740,
      "model utilizes": 20856,
      "embedding based": 9128,
      "retrieval mechanism": 28744,
      "inference validate": 14823,
      "qualitative evaluations": 26933,
      "surpasses baseline": 31738,
      "evaluation furthermore": 9954,
      "human expert": 13815,
      "expert evaluation": 10508,
      "evaluation reveals": 10003,
      "responses multiple": 28502,
      "popular opensource": 25133,
      "opensource demos": 23499,
      "gpt4 surpassing": 13122,
      "llms increasingly": 18711,
      "increasingly integrated": 14639,
      "llms comprehend": 18479,
      "comprehend interpret": 5335,
      "responses various": 28516,
      "findings revealed": 11251,
      "llms particularly": 18845,
      "particularly gpt4": 24347,
      "indicating potential": 14703,
      "using human": 34795,
      "scores models": 29280,
      "improvement models": 14343,
      "gpt4 achieving": 13054,
      "future studies": 12051,
      "significant implications": 29988,
      "education research": 8837,
      "research study": 28363,
      "development deployment": 7952,
      "leveraging capabilities": 17779,
      "gpt4 model": 13099,
      "tool designed": 33257,
      "designed quantify": 7741,
      "context analysis": 5882,
      "critical data": 6385,
      "study methods": 31361,
      "providing valuable": 26785,
      "educational outcomes": 8843,
      "novel dataset": 23071,
      "specific groups": 30695,
      "language field": 16077,
      "used create": 34593,
      "systems research": 31918,
      "objective develop": 23202,
      "pairs using": 23985,
      "order evaluate": 23673,
      "dataset findings": 6990,
      "systems dataset": 31893,
      "linguistic models": 18015,
      "capture complex": 4146,
      "complex contextual": 5269,
      "contextual relationships": 5953,
      "model meta": 20639,
      "advancement field": 1447,
      "improve natural": 14280,
      "70 billion": 495,
      "models obtain": 21764,
      "models openai": 21768,
      "openai chatgpt": 23433,
      "code research": 4814,
      "research commercial": 28298,
      "addressing challenge": 1369,
      "approach explore": 2276,
      "ensure highquality": 9604,
      "underrepresented language": 34170,
      "models strong": 22013,
      "strong linguistic": 31180,
      "linguistic properties": 18019,
      "generalpurpose llms": 12255,
      "llms leveraging": 18755,
      "study contributes": 31310,
      "adaptation strategies": 1192,
      "introducing novel": 15553,
      "publications automatic": 26847,
      "important type": 14215,
      "type information": 34053,
      "covered existing": 6321,
      "falcon vicuna": 10943,
      "models develop": 21151,
      "extraction approach": 10763,
      "approach achieves": 2228,
      "achieves improvement": 1050,
      "approach leveraging": 2310,
      "output structured": 23882,
      "information large": 14877,
      "text makes": 32910,
      "text previous": 32920,
      "previous efforts": 25868,
      "window models": 35612,
      "limited address": 17938,
      "analyze effectiveness": 1992,
      "data simply": 6867,
      "studies propose": 31279,
      "adding original": 1233,
      "enhance effectiveness": 9511,
      "effectiveness data": 8941,
      "reasoning knowledge": 27413,
      "advancements pretraining": 1474,
      "techniques models": 32652,
      "manually designed": 19570,
      "prompts work": 26447,
      "work evaluate": 35699,
      "graphs kgs": 13239,
      "llms constrained": 18488,
      "grounded kg": 13290,
      "reasoning agent": 27378,
      "textual environment": 33028,
      "state information": 30910,
      "information reasoning": 14901,
      "policy gradient": 25080,
      "gradient reinforcement": 13190,
      "algorithm model": 1710,
      "additionally conduct": 1275,
      "dataset experimental": 6986,
      "percentage points": 24460,
      "performance rate": 24731,
      "models exploring": 21242,
      "log probability": 19152,
      "increase compute": 14595,
      "inner products": 14989,
      "base methods": 3125,
      "llama7b llama13b": 18233,
      "understanding mechanism": 34250,
      "code github": 4765,
      "instruction set": 15176,
      "building llms": 3925,
      "llms languages": 18742,
      "llms understanding": 19018,
      "context problem": 5907,
      "languages need": 16901,
      "instruction sets": 15177,
      "need llms": 22637,
      "indic languages": 14685,
      "provide generative": 26704,
      "presents approach": 25573,
      "knowledge data": 15827,
      "model tailored": 20823,
      "proposed work": 26623,
      "work help": 35716,
      "set llm": 29694,
      "chinese chat": 4542,
      "sparse mixtureofexperts": 30617,
      "research demonstrated": 28301,
      "refining large": 27786,
      "machinegenerated instructionfollowing": 19369,
      "empowers models": 9273,
      "instructions paper": 15263,
      "paper systematically": 24139,
      "chinese instructionfollowing": 4544,
      "enhancing chinese": 9558,
      "conversational capabilities": 6096,
      "widely recognized": 35575,
      "effectiveness instruction": 8949,
      "finetuning sparse": 11530,
      "marking significant": 19601,
      "significant breakthrough": 29963,
      "architecture code": 2436,
      "digital forensics": 8190,
      "local large": 19129,
      "valuable tools": 35021,
      "studies explored": 31268,
      "potential chatgpt": 25249,
      "extent llms": 10722,
      "remains unresolved": 28023,
      "report evaluate": 28115,
      "evaluate strengths": 9866,
      "limitations llms": 17926,
      "different parts": 8119,
      "using case": 34747,
      "tasks inspired": 32375,
      "inspired recent": 15099,
      "present method": 25539,
      "llms called": 18455,
      "methods use": 20106,
      "available apache": 2965,
      "20 license": 211,
      "llm field": 18305,
      "era marked": 9703,
      "topic modeling": 33285,
      "modeling framework": 20893,
      "framework improve": 11865,
      "crafting prompts": 6343,
      "guide llm": 13348,
      "generation translation": 12626,
      "lexical semantic": 17801,
      "similarity generated": 30128,
      "aims reduce": 1672,
      "number topics": 23166,
      "validated effective": 34999,
      "novel insights": 23090,
      "research areas": 28294,
      "datasets specialized": 7173,
      "evaluate impact": 9840,
      "adaptation results": 1190,
      "incontext learners": 14547,
      "learners large": 17526,
      "factuality llms": 10892,
      "llms natural": 18810,
      "understanding question": 34263,
      "underexplored previous": 34143,
      "learning research": 17645,
      "models adhere": 20964,
      "users specific": 34703,
      "instructions quality": 15267,
      "explored use": 10610,
      "llms incontext": 18708,
      "inference stage": 14810,
      "effective framework": 8871,
      "framework enhances": 11849,
      "reliability llms": 27951,
      "outofdistribution data": 23747,
      "discriminative models": 8287,
      "method enhanced": 19913,
      "versions llama": 35236,
      "resources including": 28437,
      "prompts model": 26432,
      "tasks empirical": 32307,
      "analysis sheds": 1962,
      "llms highlights": 18674,
      "highlights potential": 13655,
      "methodology fostering": 19993,
      "models goal": 21321,
      "various scales": 35141,
      "scales large": 29153,
      "enhancing user": 9575,
      "behaviors different": 3326,
      "prompts extensive": 26418,
      "instructions prompts": 15265,
      "provide better": 26687,
      "guide researchers": 13351,
      "models project": 21859,
      "project page": 26245,
      "page available": 23965,
      "space recent": 30581,
      "work high": 35717,
      "mutual information": 22472,
      "information learned": 14882,
      "causal effects": 4238,
      "findings mere": 11242,
      "study focused": 31332,
      "discovered potential": 8270,
      "representational similarity": 28154,
      "similarity analysis": 30127,
      "experiments showed": 10480,
      "suggested llms": 31581,
      "trend large": 33887,
      "models involve": 21407,
      "method constructing": 19893,
      "strong model": 31182,
      "llms rarely": 18888,
      "problem based": 25987,
      "models usually": 22106,
      "function introduced": 11962,
      "used enhance": 34597,
      "approach significantly": 2337,
      "effective enhancing": 8868,
      "carefully designed": 4178,
      "new efficient": 22797,
      "efficient model": 9050,
      "conducted using": 5643,
      "dataset training": 7049,
      "accuracy efficiency": 873,
      "llm named": 18340,
      "practical application": 25360,
      "languagebased reasoning": 16857,
      "researchers develop": 28373,
      "performance hand": 24621,
      "scenarios require": 29217,
      "possibility leveraging": 25208,
      "generate plans": 12312,
      "novel hybrid": 23087,
      "complex scenarios": 5292,
      "rulebased approach": 29008,
      "performance outperforming": 24706,
      "outperforming existing": 23796,
      "metrics code": 20132,
      "evaluation need": 9981,
      "annotation tasks": 2027,
      "using open": 34863,
      "use open": 34557,
      "evaluates performance": 9890,
      "different prompting": 8127,
      "strategies models": 31111,
      "indicate need": 14692,
      "privacy reproducibility": 25954,
      "huge success": 13781,
      "ability solve": 722,
      "solve wide": 30499,
      "tasks natural": 32427,
      "applications foster": 2154,
      "domain using": 8603,
      "nlp techniques": 22965,
      "extraction document": 10766,
      "document classification": 8502,
      "work initial": 35722,
      "initial step": 14966,
      "straightforward approach": 31092,
      "30 billion": 306,
      "curated extensive": 6470,
      "data consisting": 6655,
      "knowledge answer": 15811,
      "including data": 14469,
      "collection data": 4927,
      "pretraining sft": 25838,
      "evaluation opensource": 9983,
      "opensource data": 23496,
      "exhibited impressive": 10235,
      "capabilities unclear": 4074,
      "various instructions": 35101,
      "llm alignment": 18267,
      "formulate specialized": 11771,
      "set instructions": 29690,
      "behavior llms": 3319,
      "llms address": 18426,
      "systematically comprehensively": 31877,
      "comprehensively evaluate": 5398,
      "llms responses": 18917,
      "instructions various": 15281,
      "various constraints": 35082,
      "instructions test": 15276,
      "test suite": 32790,
      "diverse forms": 8429,
      "different existing": 8076,
      "existing studies": 10316,
      "time provide": 33138,
      "provide extensive": 26699,
      "gap opensource": 12100,
      "opensource commercial": 23493,
      "research improving": 28326,
      "instructions data": 15248,
      "errors large": 9724,
      "extensive knowledge": 10709,
      "finetuning despite": 11388,
      "methods evaluating": 20029,
      "limited test": 17969,
      "testing framework": 32803,
      "factual inaccuracies": 10884,
      "llms framework": 18620,
      "approach generates": 2288,
      "types questions": 34070,
      "questions models": 27125,
      "llms factual": 18600,
      "factual accuracy": 10877,
      "accuracy increase": 892,
      "data results": 6842,
      "research endeavors": 28311,
      "study language": 31351,
      "language capability": 16047,
      "recent times": 27564,
      "mainstream llms": 19410,
      "question conduct": 27063,
      "analyze impact": 1994,
      "key factors": 15766,
      "vocabulary extension": 35382,
      "pretraining instruction": 25801,
      "accurately assess": 932,
      "assess models": 2599,
      "testing benchmarks": 32801,
      "response quality": 28480,
      "instruction tasks": 15179,
      "furthermore experimental": 12000,
      "exhibit similar": 10228,
      "community developing": 5058,
      "presents development": 25579,
      "llama2 language": 18181,
      "continuing pretraining": 5993,
      "research articles": 28295,
      "science domain": 29241,
      "methodology involves": 19994,
      "initial pretraining": 14961,
      "pretraining phase": 25831,
      "instructiontuning process": 15309,
      "process refine": 26082,
      "designed assist": 7723,
      "responses queries": 28507,
      "field materials": 11141,
      "trained checkpoints": 33384,
      "13b chat": 124,
      "effectiveness limited": 8955,
      "specialized areas": 30666,
      "15 million": 142,
      "scholarly articles": 29237,
      "specific datasets": 30687,
      "improves understanding": 14400,
      "plays crucial": 25026,
      "crucial role": 6446,
      "verifying accuracy": 35221,
      "development models": 7962,
      "community resources": 5060,
      "datasets explore": 7113,
      "enhancing llm": 9565,
      "achieve notable": 976,
      "notable improvements": 23026,
      "superior reasoning": 31656,
      "suggest continual": 31567,
      "performance specialized": 24762,
      "present extension": 25530,
      "7b llama": 535,
      "conversational dataset": 6098,
      "comprehensive quantitative": 5389,
      "ai tool": 1621,
      "era generative": 9695,
      "daily lives": 6579,
      "lives recent": 18057,
      "hold immense": 13734,
      "applications generative": 2155,
      "complex challenge": 5267,
      "identify critical": 14005,
      "challenges including": 4351,
      "high resource": 13581,
      "models prompt": 21863,
      "ondevice inference": 23340,
      "finetuning federated": 11403,
      "discuss current": 8292,
      "opensource small": 23543,
      "model present": 20719,
      "achieving better": 1086,
      "efficiency despite": 9001,
      "despite relatively": 7808,
      "demonstrates remarkable": 7564,
      "performance series": 24748,
      "llama codellama": 18088,
      "method llms": 19943,
      "llms expansion": 18584,
      "transformer blocks": 33712,
      "using new": 34857,
      "new corpus": 22788,
      "effectively improving": 8922,
      "improving models": 14416,
      "knowledge catastrophic": 15822,
      "corpus code": 6175,
      "model initialized": 20583,
      "existing open": 10303,
      "intelligent agent": 15363,
      "natural programming": 22578,
      "laying solid": 17452,
      "solid foundation": 30466,
      "developing advanced": 7940,
      "advanced language": 1423,
      "language agents": 16041,
      "effectively various": 8936,
      "rapid development": 27247,
      "development opensource": 7967,
      "llms truly": 19012,
      "study scaling": 31392,
      "facilitate scaling": 10846,
      "advancing opensource": 1494,
      "developed dataset": 7924,
      "conduct supervised": 5621,
      "sft direct": 29760,
      "models resulting": 21940,
      "resulting creation": 28553,
      "surpasses llama2": 31746,
      "exhibits superior": 10255,
      "compared gpt35": 5137,
      "gpt35 mixtral": 13029,
      "experts introduce": 10519,
      "mistral 7b": 20228,
      "token layer": 33193,
      "process current": 26054,
      "parameters inference": 24257,
      "trained context": 33386,
      "context size": 5919,
      "32k tokens": 328,
      "evaluated benchmarks": 9874,
      "benchmarks particular": 3466,
      "mathematics code": 19687,
      "generation multilingual": 12558,
      "finetuned follow": 11310,
      "instruct surpasses": 15130,
      "surpasses gpt35": 31745,
      "gpt35 turbo": 13036,
      "pro llama": 25962,
      "chat model": 4443,
      "model human": 20566,
      "base instruct": 3117,
      "instruct models": 15129,
      "models released": 21918,
      "released apache": 27919,
      "hierarchical spatial": 13543,
      "questions designed": 27107,
      "challenge llms": 4318,
      "llms scenarios": 18928,
      "scenarios potentially": 29215,
      "followed gpt35": 11683,
      "models showed": 21967,
      "significantly reduced": 30082,
      "models identified": 21366,
      "discuss potential": 8295,
      "textbased data": 32972,
      "multimodal large": 22351,
      "llms multimodal": 18805,
      "models mllms": 21731,
      "mllms shown": 20300,
      "shown excellent": 29873,
      "domainspecific benchmarks": 8646,
      "performance mllms": 24679,
      "modern society": 22172,
      "reliably perform": 27958,
      "tasks address": 32236,
      "understanding applying": 34210,
      "results models": 28647,
      "serve foundation": 29647,
      "foundation future": 11791,
      "models mental": 21716,
      "health challenges": 13485,
      "models facilitated": 21255,
      "applications significant": 2177,
      "significant research": 30020,
      "understanding enhancing": 34219,
      "models domain": 21168,
      "question involves": 27071,
      "capacity large": 4126,
      "models comprehend": 21077,
      "study presents": 31377,
      "models addressing": 20962,
      "addressing gap": 1371,
      "performance llama2": 24661,
      "llama2 chatgpt": 18168,
      "individual preferences": 14719,
      "shown llms": 29900,
      "finetuned generate": 11316,
      "reward functions": 28858,
      "generation improve": 12519,
      "demonstrate great": 7460,
      "llms suffering": 18980,
      "propose inferencetime": 26519,
      "help llms": 13507,
      "llms decode": 18507,
      "information theory": 14918,
      "tokens predicted": 33242,
      "lower probabilities": 19290,
      "analysis shows": 1966,
      "closely related": 4693,
      "original context": 23702,
      "forcing model": 11720,
      "requiring additional": 28268,
      "additional data": 1254,
      "improvements achieved": 14354,
      "llama27b mistral7b": 18208,
      "datasets included": 7131,
      "phenomenon known": 24923,
      "performance little": 24659,
      "understanding potential": 34261,
      "stage pretraining": 30858,
      "text evaluation": 32854,
      "evaluation samples": 10005,
      "data investigate": 6743,
      "current llm": 6510,
      "offer new": 23290,
      "new insights": 22809,
      "insights data": 15070,
      "effects language": 8981,
      "underscore need": 34175,
      "mixtureofexperts language": 20285,
      "costs scaling": 6274,
      "knowledge response": 15903,
      "response propose": 28479,
      "allowing flexible": 1805,
      "15 times": 144,
      "parameters set": 24286,
      "16b parameters": 165,
      "architecture performance": 2450,
      "llms hold": 18676,
      "investigates potential": 15606,
      "sequential data": 29625,
      "complex data": 5271,
      "llm study": 18371,
      "study focuses": 31333,
      "scenarios findings": 29206,
      "demonstrate models": 7476,
      "models proficiency": 21858,
      "token length": 33194,
      "length limitations": 17707,
      "open new": 23410,
      "new avenues": 22778,
      "models search": 21957,
      "tuning large": 33987,
      "ir tasks": 15641,
      "promptbased methods": 26363,
      "facilitating comprehensive": 10852,
      "understanding execution": 34221,
      "gap work": 12115,
      "document understanding": 8509,
      "analyze effects": 1993,
      "fewshot demonstrations": 11100,
      "make dataset": 19465,
      "domain use": 8602,
      "acquire ability": 1115,
      "answering training": 2079,
      "domainspecific questions": 8655,
      "users queries": 34701,
      "frequently asked": 11926,
      "asked questions": 2557,
      "embedding model": 9130,
      "model terms": 20827,
      "generate answer": 12262,
      "llm optimize": 18341,
      "model external": 20508,
      "policy optimize": 25089,
      "using policy": 34873,
      "multiple training": 22426,
      "significant cost": 29975,
      "improved accuracy": 14305,
      "rl approach": 28900,
      "approach generic": 2290,
      "existing rag": 10312,
      "model commonsense": 20430,
      "intermediate steps": 15433,
      "accomplish task": 841,
      "series modifications": 29642,
      "effectively reason": 8927,
      "understand inputs": 34192,
      "presents challenging": 25576,
      "procedural text": 26043,
      "data natural": 6780,
      "evidence supporting": 10066,
      "using multiple": 34848,
      "tasks sequencetosequence": 32496,
      "metrics particular": 20146,
      "crosstask knowledge": 6428,
      "reusing data": 28788,
      "data different": 6676,
      "lead higher": 17465,
      "optimization strategy": 23635,
      "does yield": 8543,
      "yield significant": 35913,
      "improvement model": 14342,
      "t5small model": 31979,
      "model synthetic": 20820,
      "account model": 851,
      "size decreases": 30245,
      "multilingual evaluation": 22307,
      "models parameter": 21800,
      "viable solution": 35242,
      "solution improving": 30475,
      "llms requiring": 18914,
      "shown large": 29893,
      "models equitable": 21211,
      "work finetune": 35710,
      "finetune llama27b": 11290,
      "tuning datasets": 33972,
      "tasks covering": 32280,
      "various parameters": 35132,
      "effects downstream": 8978,
      "bridges gap": 3872,
      "ones english": 23346,
      "performance lowresource": 24668,
      "degrading performance": 7385,
      "importance data": 14186,
      "pretraining monolingual": 25823,
      "paper conduct": 24023,
      "comprehensive study": 5392,
      "performance suite": 24774,
      "suite stateoftheart": 31601,
      "performance leading": 24652,
      "suggests pretraining": 31591,
      "performance surpassing": 24776,
      "training better": 33445,
      "performance reasoning": 24733,
      "tasks tend": 32524,
      "data languages": 6748,
      "languages train": 16916,
      "incurs high": 14664,
      "translated data": 33813,
      "reasoning questions": 27443,
      "english finetuning": 9479,
      "alignment makes": 1770,
      "project available": 26244,
      "excel tasks": 10154,
      "important measure": 14204,
      "reflect models": 27791,
      "models behavior": 21011,
      "measure called": 19733,
      "llms inference": 18718,
      "successfully applied": 31536,
      "falcon 40b": 10938,
      "numerical calculations": 23178,
      "scientific reasoning": 29256,
      "central approach": 4267,
      "approach novel": 2319,
      "address data": 1320,
      "stepbystep reasoning": 31055,
      "dataset encompassing": 6980,
      "largerscale models": 17341,
      "capabilities base": 4002,
      "model makes": 20634,
      "wider research": 35586,
      "framework finetuning": 11855,
      "finetuning pipelines": 11484,
      "llms retrievalaugmented": 18920,
      "rag augments": 27156,
      "external data": 10727,
      "understood paper": 34281,
      "propose pipeline": 26559,
      "pipeline finetuning": 24967,
      "including llama213b": 14502,
      "extracting information": 10757,
      "generating questions": 12440,
      "answers using": 2087,
      "metrics assess": 20131,
      "pipeline conduct": 24965,
      "conduct indepth": 5612,
      "indepth study": 14678,
      "study potentially": 31375,
      "disruptive application": 8321,
      "dataset generation": 6997,
      "finetuning accuracy": 11367,
      "demonstrate finetuned": 7455,
      "model leverages": 20612,
      "specific questions": 30712,
      "questions increasing": 27114,
      "47 72": 403,
      "results point": 28655,
      "paving way": 24421,
      "llms industrial": 18717,
      "despite general": 7777,
      "models consistently": 21100,
      "consistently benefit": 5746,
      "achieve desired": 960,
      "blackbox lms": 3755,
      "prediction output": 25431,
      "smaller lm": 30379,
      "larger scale": 17336,
      "pretraining experiments": 25796,
      "knowledge reasoning": 15897,
      "reasoning safety": 27450,
      "tuned models": 33962,
      "models factual": 21258,
      "knowledge demonstrate": 15828,
      "demonstrate generality": 7458,
      "finetuning questionanswering": 11501,
      "promise using": 26278,
      "novel adaptive": 23052,
      "learning proposed": 17638,
      "tasks train": 32534,
      "validation performance": 35006,
      "optimization framework": 23626,
      "finally provide": 11203,
      "analysis interpolation": 1931,
      "aimed enhancing": 1652,
      "addresses key": 1365,
      "offers solution": 23313,
      "approach includes": 2299,
      "scenarios framework": 29207,
      "dialogues accurately": 8024,
      "boosting user": 3826,
      "marks significant": 19605,
      "models synthesize": 22036,
      "300b tokens": 311,
      "tokens model": 33240,
      "tokens included": 33233,
      "pretrained llama2": 25704,
      "domainspecific dataset": 8649,
      "finetuned highquality": 11319,
      "retrieval strategy": 28755,
      "perform comparably": 24474,
      "scale large": 29137,
      "number languages": 23149,
      "different research": 8133,
      "focusing different": 11671,
      "answer propose": 2048,
      "performs par": 24850,
      "results general": 28618,
      "benchmarks models": 3464,
      "trained evaluated": 33397,
      "train reward": 33372,
      "performance level": 24654,
      "learn improve": 17508,
      "prompting provide": 26393,
      "dpo training": 8708,
      "training does": 33502,
      "does instruction": 8531,
      "following ability": 11687,
      "provide highquality": 26705,
      "iterations approach": 15683,
      "yields model": 35925,
      "alpacaeval 20": 1838,
      "work opens": 35739,
      "possibility models": 25209,
      "semantic content": 29453,
      "tools like": 33271,
      "like wikipedia": 17897,
      "product descriptions": 26168,
      "produce structured": 26154,
      "offering practical": 23299,
      "practical solution": 25372,
      "focus improving": 11649,
      "models applied": 20983,
      "instructions results": 15271,
      "llms task": 18995,
      "generation generating": 12510,
      "generating coherent": 12414,
      "data avoid": 6619,
      "novel structured": 23111,
      "public apis": 26831,
      "mistral zephyr": 20235,
      "generate fluent": 12278,
      "standard data": 30871,
      "data formats": 6718,
      "semantic accuracy": 29446,
      "major issue": 19442,
      "knowledge fusion": 15853,
      "fusion large": 12022,
      "models distinct": 21162,
      "varying architectures": 35168,
      "introduce notion": 15523,
      "combining capabilities": 4962,
      "capabilities existing": 4015,
      "single llm": 30211,
      "source llms": 30566,
      "unique strengths": 34364,
      "llms improve": 18686,
      "performance target": 24778,
      "capabilities reasoning": 4066,
      "generation code": 12472,
      "weights data": 35504,
      "data public": 6818,
      "simple framework": 30149,
      "english nonenglish": 9488,
      "llama trained": 18147,
      "especially pronounced": 9746,
      "generation address": 12452,
      "framework designed": 11838,
      "linguistic units": 18024,
      "tailored target": 32019,
      "reducing number": 27758,
      "compared standard": 5170,
      "standard decoding": 30875,
      "maintaining performance": 19430,
      "benchmark language": 3395,
      "target word": 32061,
      "word context": 35634,
      "context sentence": 5916,
      "designed assess": 7722,
      "generate appropriate": 12264,
      "propose models": 26532,
      "models automatically": 20999,
      "finetuned taskspecific": 11358,
      "training memory": 33561,
      "updating small": 34479,
      "lm parameters": 19059,
      "does improve": 8530,
      "structured pruning": 31228,
      "improves lm": 14383,
      "memory time": 19831,
      "efficiency introduce": 9004,
      "parameters lms": 24271,
      "tuning parameters": 34003,
      "efficiency compared": 8998,
      "performance pruning": 24727,
      "70 parameters": 497,
      "lms finetuning": 19085,
      "opensource multilingual": 23537,
      "study introduce": 31342,
      "diverse corpus": 8420,
      "tokens sourced": 33246,
      "english chinese": 9476,
      "specific use": 30723,
      "performance broad": 24537,
      "tasks make": 32416,
      "associated code": 2643,
      "inspire future": 15091,
      "research practical": 28342,
      "commonly employ": 5024,
      "high memory": 13574,
      "decoding models": 7278,
      "plugin module": 25058,
      "memory costs": 19811,
      "applying proposed": 2222,
      "mtbench benchmark": 22260,
      "surpasses stateoftheart": 31750,
      "explainability large": 10526,
      "llms critical": 18498,
      "aspect natural": 2565,
      "processing llms": 26111,
      "increasingly integral": 14638,
      "blackbox nature": 3757,
      "ethical use": 9807,
      "focus primarily": 11653,
      "transformerbased llms": 33757,
      "challenges scale": 4375,
      "terms existing": 32745,
      "model editing": 20478,
      "control generation": 6050,
      "advantages limitations": 1507,
      "span corruption": 30590,
      "models known": 21412,
      "resource intensive": 28412,
      "sequences paper": 29615,
      "empirically effectiveness": 9248,
      "twostage pretraining": 34045,
      "analysis case": 1911,
      "architectures t5": 2473,
      "pretraining enabling": 25795,
      "40 reduction": 374,
      "performance research": 24738,
      "ability llm": 696,
      "emerged powerful": 9163,
      "detection llms": 7869,
      "assess capabilities": 2588,
      "collected different": 4923,
      "performance detecting": 24566,
      "content aigc": 5850,
      "test ability": 32758,
      "ability distinguish": 662,
      "llms identify": 18682,
      "generated ai": 12341,
      "long way": 19190,
      "fundamental task": 11982,
      "task information": 32139,
      "information systems": 14915,
      "heavily rely": 13494,
      "collecting annotating": 4925,
      "newly emerging": 22870,
      "timeconsuming laborintensive": 33149,
      "recent advanced": 27487,
      "inspiring explore": 15104,
      "explore alternative": 10571,
      "propose zeroshot": 26586,
      "data retrieval": 6844,
      "prompt guide": 26328,
      "chatgpt generate": 4472,
      "data step": 6881,
      "step improve": 31047,
      "knowledge leveraging": 15878,
      "public datasets": 26835,
      "results illustrate": 28626,
      "model integration": 20591,
      "integration paper": 15345,
      "knowledge multiple": 15883,
      "core framework": 6151,
      "task execution": 32116,
      "query generation": 27027,
      "generate training": 12335,
      "mmlu benchmark": 20314,
      "13b 34b": 121,
      "demonstrate new": 7477,
      "mixtral model": 20274,
      "illustrate potential": 14047,
      "architecture creating": 2438,
      "benchmark chinese": 3357,
      "chinese multimodal": 4548,
      "reasoning multimodal": 27425,
      "language modelsmllms": 16763,
      "progress demonstrated": 26209,
      "comprehension reasoning": 5349,
      "challenge current": 4308,
      "current multimodal": 6517,
      "multimodal benchmarks": 22337,
      "benchmark multimodal": 3400,
      "greater challenges": 13268,
      "opensource mllms": 23526,
      "mllms gpt4v": 20297,
      "gpt4v geminipro": 13136,
      "poses significant": 25173,
      "extreme compression": 10791,
      "llama advancing": 18072,
      "size poses": 30271,
      "huge training": 13782,
      "compression methods": 5420,
      "network quantization": 22701,
      "keeping number": 15746,
      "compression approach": 5412,
      "benchmark demonstrate": 3377,
      "addresses challenge": 1362,
      "challenge extending": 4312,
      "extending large": 10663,
      "propose approach": 26496,
      "shared tokens": 29786,
      "tokens english": 33224,
      "alignment approach": 1754,
      "text reduces": 32930,
      "mt tasks": 22251,
      "closer alignment": 4695,
      "leveraging power": 17791,
      "english llms": 9485,
      "encoderdecoder language": 9366,
      "model enhanced": 20491,
      "advances natural": 1483,
      "languages work": 16922,
      "pretrained encoderdecoder": 25640,
      "based unified": 3235,
      "tasks understanding": 32537,
      "outperforms multilingual": 23837,
      "rows columns": 28998,
      "processing use": 26134,
      "comes substantial": 4975,
      "costs terms": 6275,
      "resource constraints": 28409,
      "challenges need": 4363,
      "need additional": 22620,
      "hardware paper": 13431,
      "reducing embedding": 27746,
      "zeroshot task": 35997,
      "models run": 21950,
      "gpus reduce": 13180,
      "40gb a100": 381,
      "hope inspire": 13754,
      "reduce memory": 27718,
      "memory computation": 19806,
      "makes inference": 19491,
      "timeconsuming paper": 33150,
      "key observations": 15781,
      "token level": 33195,
      "inherent uncertainty": 14951,
      "based insights": 3179,
      "insights introduce": 15078,
      "simple highly": 30152,
      "enabling precise": 9328,
      "conducted comprehensive": 5631,
      "tasks dialogue": 32296,
      "dialogue code": 8011,
      "llama2chat 70b": 18217,
      "distribution generated": 8393,
      "tool use": 33260,
      "data analysis": 6596,
      "finance large": 11210,
      "error propagation": 9714,
      "mitigate limitations": 20255,
      "certain reasoning": 4278,
      "task instead": 32141,
      "llms inherent": 18721,
      "inherent abilities": 14946,
      "using financial": 34778,
      "financial domain": 11218,
      "questionanswering datasets": 27084,
      "apply supervised": 2213,
      "finetuning llama2": 11443,
      "task solver": 32193,
      "right tool": 28880,
      "demonstrates improvement": 7561,
      "results best": 28576,
      "augmentation language": 2800,
      "models finance": 21271,
      "finance domain": 11209,
      "human learning": 13840,
      "continuous feedback": 5996,
      "inspired paper": 15097,
      "novel teacherstudent": 23115,
      "agent provides": 1563,
      "feedback forms": 11061,
      "posed questions": 25164,
      "reasoning testbed": 27460,
      "training llama2": 33552,
      "llama2 data": 18170,
      "training curriculum": 33464,
      "learning robustness": 17648,
      "finetuning variety": 11553,
      "imagetext instruction": 14093,
      "versatile multimodal": 35224,
      "model mllm": 20643,
      "different capabilities": 8055,
      "distinct domains": 8365,
      "tasks specific": 32512,
      "propose apply": 26495,
      "lowrank adaption": 19304,
      "adaption lora": 1219,
      "lora method": 19232,
      "set lora": 29695,
      "original lora": 23713,
      "effectively mitigates": 8925,
      "multiple distinct": 22388,
      "various configurations": 35081,
      "achieves consistent": 1045,
      "consistent performance": 5741,
      "claim verification": 4577,
      "automated factchecking": 2861,
      "evidence work": 10067,
      "limited data": 17948,
      "supervision propose": 31698,
      "pioneering approach": 24960,
      "novel semantic": 23106,
      "leverages unlabelled": 17775,
      "data annotations": 6598,
      "computing resources": 5518,
      "improvements sota": 14364,
      "sota baselines": 30532,
      "factchecking datasets": 10859,
      "methods neural": 20069,
      "explore challenges": 10578,
      "computational storage": 5482,
      "tokens following": 33228,
      "benchmarks additionally": 3430,
      "additionally release": 1301,
      "trained supervised": 33430,
      "finetuning followed": 11406,
      "generation text": 12621,
      "generate humanreadable": 12290,
      "used text": 34629,
      "major computational": 19441,
      "generation unlike": 12628,
      "tokens parallel": 33241,
      "paper proposed": 24119,
      "architecture named": 2448,
      "architecture utilizes": 2458,
      "integrating multiple": 15338,
      "optimized data": 23645,
      "accelerates endtoend": 797,
      "endtoend inference": 9434,
      "generation furthermore": 12508,
      "furthermore validate": 12015,
      "input size": 15030,
      "achieves maximum": 1051,
      "times speedup": 33166,
      "llms evaluating": 18572,
      "aligning large": 1745,
      "models news": 21752,
      "challenge especially": 4311,
      "llms enhance": 18564,
      "llms evaluation": 18573,
      "metrics account": 20130,
      "biases generated": 3672,
      "unclear models": 34126,
      "tasks specifically": 32513,
      "impacts models": 14147,
      "news articles": 22876,
      "online news": 23366,
      "models mistral7b": 21727,
      "gpt3 using": 13009,
      "sample baseline": 29064,
      "instructionbased models": 15215,
      "comparison finetuned": 5195,
      "smaller llms": 30378,
      "better reflect": 3623,
      "open corpus": 23392,
      "tokens language": 33236,
      "models critical": 21115,
      "models rarely": 21894,
      "data open": 6787,
      "datasets trained": 7183,
      "result challenging": 28543,
      "english corpus": 9477,
      "corpus built": 6174,
      "built diverse": 3932,
      "web content": 35475,
      "work report": 35773,
      "including design": 14470,
      "design principles": 7714,
      "content quality": 5870,
      "data architectures": 6602,
      "biases potential": 3681,
      "access powerful": 825,
      "open lms": 23407,
      "framework build": 11833,
      "code release": 4805,
      "framework including": 11867,
      "new wave": 22865,
      "tasks explicitly": 32324,
      "requires substantial": 28263,
      "resources paper": 28442,
      "compact llms": 5065,
      "task realworld": 32184,
      "llama2 gpt35": 18178,
      "observe smaller": 23234,
      "summarization datasets": 31610,
      "parameters performs": 24275,
      "better zeroshot": 3637,
      "7b 70b": 530,
      "like flant5": 17862,
      "robustness data": 28945,
      "data compression": 6648,
      "compression existing": 5414,
      "models face": 21253,
      "compression based": 5413,
      "models predictive": 21833,
      "specifically collect": 30729,
      "data spanning": 6876,
      "data cutoff": 6668,
      "compression performance": 5422,
      "performance testing": 24782,
      "unseen data": 34434,
      "measure robustness": 19737,
      "experiments test": 10491,
      "representative large": 28181,
      "sources including": 30575,
      "wikipedia news": 35604,
      "multimodal data": 22340,
      "models mistral": 21726,
      "mistral llama2": 20232,
      "llama2 demonstrate": 18171,
      "good balance": 12818,
      "balance performance": 3080,
      "performance robustness": 24743,
      "struggle generalize": 31239,
      "capable performing": 4116,
      "difficult deploy": 8170,
      "gpt4 smaller": 13119,
      "using zeroshot": 34942,
      "prompting generate": 26374,
      "near 100": 22599,
      "100 success": 39,
      "previous methods": 25869,
      "gpt4 finetune": 13076,
      "sizes gpt2": 30299,
      "holdout test": 13738,
      "set gpt2": 29688,
      "achieves 90": 1026,
      "90 success": 591,
      "success gpt4": 31510,
      "task evaluating": 32115,
      "evaluating quality": 9912,
      "distilled models": 8354,
      "zeroshot classifier": 35963,
      "classifier achieves": 4624,
      "smaller draft": 30375,
      "draft model": 8711,
      "target llm": 32054,
      "depends choice": 7623,
      "provides high": 26755,
      "generated token": 12396,
      "llm achieve": 18262,
      "throughput experiments": 33099,
      "understand phenomenon": 34198,
      "factors affect": 10867,
      "analytical model": 1983,
      "higher throughput": 13608,
      "draft models": 8712,
      "forms foundation": 11767,
      "work directly": 35695,
      "data users": 6905,
      "users device": 34685,
      "utilizing large": 34973,
      "ondevice deployment": 23339,
      "novel lightweight": 23092,
      "lightweight framework": 17839,
      "framework enhanced": 11848,
      "related text": 27858,
      "effectiveness leveraging": 8954,
      "significantly achieves": 30030,
      "new sota": 22845,
      "improvement bleu": 14334,
      "bleu meteor": 3765,
      "train release": 33371,
      "moe llms": 22197,
      "llms ranging": 18887,
      "llms offer": 18824,
      "llms highlighting": 18673,
      "llm development": 18289,
      "contribution study": 6038,
      "models predominantly": 21834,
      "based token": 3231,
      "token ids": 33192,
      "early pretraining": 8779,
      "llms simulate": 18954,
      "gpt4 mixtral": 13098,
      "humancomputer interaction": 13893,
      "human voting": 13876,
      "baseline human": 3248,
      "inherent biases": 14947,
      "humans llms": 13927,
      "llms observed": 18823,
      "diverse preferences": 8448,
      "llms lead": 18747,
      "underscoring need": 34185,
      "integration llms": 15344,
      "use generative": 34532,
      "ai chatbots": 1603,
      "learning paper": 17625,
      "includes investigation": 14451,
      "using blooms": 34742,
      "initial results": 14963,
      "lays foundation": 17457,
      "paper concludes": 24022,
      "faster lighter": 11001,
      "way forward": 35432,
      "computational memory": 5470,
      "requirements inference": 28238,
      "methods aim": 20005,
      "enhance llm": 9516,
      "overview methods": 23955,
      "providing practical": 26782,
      "llm deployment": 18286,
      "unified setting": 34336,
      "highlights effectiveness": 13650,
      "identify current": 14006,
      "limitations discuss": 17917,
      "directions improve": 8230,
      "presented paper": 25567,
      "guardrails large": 13335,
      "crucial identify": 6441,
      "identify mitigate": 14012,
      "mitigate risks": 20259,
      "human users": 13873,
      "outputs llms": 23895,
      "position paper": 25184,
      "paper takes": 24141,
      "current opensource": 6521,
      "llama guard": 18112,
      "approach construct": 2253,
      "applications propose": 2172,
      "transformers pretrained": 33793,
      "finetuning parameters": 11473,
      "parameters challenging": 24230,
      "gained popularity": 12064,
      "intermediate outputs": 15429,
      "enables efficient": 9297,
      "inference sparsityaware": 14807,
      "building insight": 3924,
      "approach utilizing": 2359,
      "lora adapter": 19227,
      "facilitate efficient": 10838,
      "experiments proposed": 10467,
      "reduce activation": 27699,
      "benchmarks respectively": 3470,
      "tasks showcase": 32499,
      "pruned models": 26804,
      "capture semantics": 4150,
      "reasoning reading": 27444,
      "opensource instructiontuned": 23505,
      "instructiontuned llama": 15292,
      "stateoftheart multitask": 30961,
      "multitask finetuned": 22445,
      "compared llama": 5145,
      "including code": 14468,
      "model dataset": 20451,
      "relatively new": 27891,
      "new decoding": 22792,
      "leverages small": 17774,
      "models reduce": 21913,
      "reduce latency": 27716,
      "frozen llm": 11938,
      "method uses": 19982,
      "models confidence": 21093,
      "scores help": 29279,
      "benchmarks demonstrate": 3434,
      "data trained": 6895,
      "models enhancing": 21207,
      "approach applicable": 2236,
      "method involves": 19936,
      "interpret context": 15454,
      "addressing inherent": 1373,
      "historical information": 13726,
      "single hidden": 30204,
      "hidden state": 13538,
      "improvement achieved": 14326,
      "increase number": 14601,
      "parameters little": 24269,
      "original number": 23715,
      "additional parameters": 1262,
      "minimal computational": 20184,
      "pretraining resulting": 25834,
      "linear computational": 17987,
      "approach showcasing": 2336,
      "showcasing improved": 29852,
      "benchmarks code": 3432,
      "weights datasets": 35506,
      "datasets opensourced": 7158,
      "lottery ticket": 19263,
      "ticket hypothesis": 33103,
      "hypothesis posits": 13965,
      "method identify": 19929,
      "llm parameters": 18346,
      "highly effective": 13662,
      "idea use": 13985,
      "parameters finetuning": 24246,
      "theoretically prove": 33054,
      "set parameters": 29701,
      "embedding llama": 9129,
      "performance code": 24546,
      "model released": 20751,
      "released public": 27927,
      "semantic representations": 29468,
      "malaysian language": 19521,
      "specifically llama2": 30748,
      "pairs release": 23984,
      "outperforms openai": 23839,
      "rag models": 27159,
      "context notably": 5904,
      "underscore effectiveness": 34173,
      "text comprehension": 32832,
      "work studies": 35789,
      "premises important": 25496,
      "complex multihop": 5280,
      "datasets contain": 7084,
      "contain short": 5829,
      "challenges address": 4335,
      "benchmark includes": 3390,
      "includes datasets": 14450,
      "datasets nlp": 7155,
      "nlp domains": 22932,
      "extended contexts": 10660,
      "humans perform": 13929,
      "deductive reasoning": 7310,
      "obtain strong": 23254,
      "strong opensource": 31184,
      "outperforms gpt35": 23826,
      "gpt4 finally": 13075,
      "use model": 34550,
      "model filter": 20520,
      "improvement average": 14330,
      "llms proven": 18878,
      "hallucination responses": 13384,
      "responses lack": 28498,
      "intuitive solution": 15559,
      "works directly": 35812,
      "llms work": 19048,
      "propose effective": 26506,
      "effective training": 8903,
      "generate highly": 12284,
      "ensuring correctness": 9610,
      "correctness responses": 6209,
      "responses conduct": 28488,
      "conventional practices": 6078,
      "models generalizability": 21304,
      "answering task": 2076,
      "task large": 32148,
      "task artificial": 32079,
      "data previous": 6806,
      "research significantly": 28361,
      "limitations including": 17920,
      "inability capture": 14430,
      "capture contextual": 4148,
      "directly applying": 8233,
      "applying llms": 2218,
      "llms leads": 18748,
      "llm model": 18338,
      "model llama2": 20619,
      "rag architecture": 27155,
      "architecture outperforms": 2449,
      "course training": 6312,
      "degradation model": 7375,
      "quality smaller": 26979,
      "alternative framework": 1852,
      "better pretraining": 3618,
      "pretraining loss": 25818,
      "loss bert": 19242,
      "ul2 language": 34095,
      "shows better": 29923,
      "better downstream": 3599,
      "provide theoretical": 26730,
      "increasing complexity": 14619,
      "residual connections": 28390,
      "layer norm": 17427,
      "sparsity large": 30630,
      "high inference": 13569,
      "emergence activation": 9167,
      "sparsity llms": 30634,
      "llms provides": 18880,
      "natural approach": 22503,
      "reduce cost": 27707,
      "inference existing": 14775,
      "achieve introduce": 971,
      "furthermore unlike": 12013,
      "applied llms": 2195,
      "activation functions": 1140,
      "tasks outperforming": 32438,
      "outperforming stateoftheart": 23803,
      "november 2022": 23129,
      "training billions": 33446,
      "models parameters": 21803,
      "different ways": 8158,
      "including popular": 14512,
      "llama palm": 18139,
      "techniques developed": 32635,
      "augment llms": 2791,
      "used llm": 34608,
      "performance popular": 24715,
      "llms set": 18933,
      "benchmarks finally": 3441,
      "open challenges": 23386,
      "nlp attributed": 22922,
      "llm respond": 18361,
      "respond instructions": 28468,
      "finetuning ift": 11416,
      "annotated datasets": 2020,
      "datasets existing": 7106,
      "primary goal": 25922,
      "language gap": 16080,
      "create extensive": 6350,
      "date comprising": 7198,
      "resources develop": 28432,
      "develop opensource": 7921,
      "evaluation suite": 10015,
      "serves valuable": 29656,
      "framework future": 11858,
      "aim bridge": 1637,
      "agent framework": 1560,
      "emerging building": 9191,
      "data diverse": 6681,
      "heavily relies": 13493,
      "relies manual": 27966,
      "hindering potential": 13717,
      "extraction knowledge": 10768,
      "graph completion": 13218,
      "generation propose": 12579,
      "refinement module": 27782,
      "human gpt4": 13823,
      "surpass stateoftheart": 31734,
      "gpt4 10": 13050,
      "20 times": 214,
      "existing benchmark": 10264,
      "model recent": 20744,
      "multiple domains": 22389,
      "graph reasoning": 13229,
      "reasoning llm": 27419,
      "llm remains": 18357,
      "remains limited": 28004,
      "limited work": 17974,
      "llms graph": 18662,
      "reasoning performance": 27433,
      "significantly affects": 30034,
      "llms reasoning": 18890,
      "altering order": 1846,
      "order enhance": 23672,
      "assessing llms": 2611,
      "relationship llms": 27873,
      "abilities graph": 625,
      "experiments span": 10482,
      "models rlhf": 21947,
      "llm behaviors": 18276,
      "controllable inference": 6059,
      "instructing llm": 15139,
      "critiques revisions": 6403,
      "finetuning synthetic": 11540,
      "problem llms": 25995,
      "llms generative": 18639,
      "llms great": 18663,
      "potential reshape": 25293,
      "landscape social": 16027,
      "malicious actors": 19523,
      "developed llms": 7928,
      "experimental framework": 10388,
      "bot human": 3836,
      "surveyed participants": 31778,
      "human detection": 13801,
      "correctly identified": 6204,
      "impact human": 14124,
      "human perception": 13850,
      "received lot": 27481,
      "performance understanding": 24789,
      "human languages": 13839,
      "languages lowresource": 16893,
      "resources work": 28449,
      "focus enhancing": 11646,
      "model integrating": 20590,
      "model shows": 20781,
      "results different": 28599,
      "opensource dataset": 23497,
      "studies models": 31276,
      "increasingly utilized": 14649,
      "blackbox models": 3756,
      "studies sought": 31286,
      "llms previous": 18868,
      "studies provided": 31280,
      "study seeks": 31395,
      "different input": 8083,
      "input prompts": 15022,
      "prompts specifically": 26441,
      "prompts designed": 26410,
      "examine models": 10103,
      "models susceptible": 22035,
      "like humans": 17875,
      "based language": 3183,
      "lms different": 19082,
      "including llama2": 14501,
      "llama2 falcon": 18173,
      "gpt opt": 12860,
      "fine tuned": 11264,
      "using classifiers": 34750,
      "dataset results": 7034,
      "demonstrate high": 7461,
      "broader range": 3891,
      "text embedding": 32849,
      "answering question": 2070,
      "aims build": 1660,
      "despite tremendous": 7823,
      "tremendous potential": 33883,
      "texts implicit": 32994,
      "similar embeddings": 30101,
      "models abstractive": 20938,
      "improved instructionfollowing": 14313,
      "llms llama27b": 18780,
      "additionally qualitative": 1299,
      "applying different": 2217,
      "different instructions": 8085,
      "instructionfollowing capability": 15225,
      "prompts effective": 26412,
      "effective methods": 8885,
      "methods proposed": 20081,
      "constraint prompt": 5785,
      "novel connection": 23068,
      "based connection": 3145,
      "general framework": 12166,
      "harness power": 13454,
      "datasets finetuning": 7118,
      "computational demand": 5464,
      "adds new": 1377,
      "components additional": 5311,
      "introduce simple": 15532,
      "performance interesting": 24634,
      "interesting finding": 15407,
      "models enabling": 21196,
      "enabling use": 9329,
      "experiments llama2": 10459,
      "families models": 10969,
      "showcasing minimal": 29853,
      "understanding incontext": 34232,
      "processing based": 26096,
      "latent variable": 17411,
      "variable models": 35033,
      "framework framework": 11857,
      "framework introduce": 11871,
      "sense knowledge": 29506,
      "language classification": 16050,
      "effect choice": 8851,
      "possible explain": 25213,
      "generalization tasks": 12227,
      "tasks unseen": 32541,
      "unseen language": 34439,
      "framework explain": 11854,
      "work english": 35697,
      "multilingual transformers": 22334,
      "importance understanding": 14193,
      "layer layer": 17426,
      "layer transformers": 17433,
      "map input": 19577,
      "input embedding": 15009,
      "output embedding": 23866,
      "probabilities computed": 25969,
      "distinct phases": 8368,
      "far away": 10987,
      "semantically correct": 29483,
      "correct token": 6195,
      "higher probability": 13602,
      "input language": 15014,
      "conceptual model": 5535,
      "input space": 15031,
      "concept space": 5528,
      "output space": 23881,
      "evidence suggests": 10063,
      "languages important": 16877,
      "precision recall": 25393,
      "recall assess": 27469,
      "quality diversity": 26954,
      "novel evaluation": 23077,
      "framework large": 11874,
      "image generation": 14067,
      "diversity generated": 8478,
      "insights performance": 15080,
      "performance openended": 24702,
      "tasks adequately": 32238,
      "generated samples": 12387,
      "finetuned human": 11320,
      "work extends": 35709,
      "offering insights": 23298,
      "capabilities challenges": 4004,
      "challenges faced": 4343,
      "framework inspired": 11869,
      "answering cqa": 2058,
      "35 llama": 341,
      "llama experiments": 18096,
      "additional analyses": 1250,
      "significantly correlated": 30042,
      "opening opportunities": 23469,
      "opportunities future": 23582,
      "future development": 12029,
      "verification tools": 35211,
      "recurrent memory": 27679,
      "evaluate different": 9825,
      "different approaches": 8048,
      "evaluation includes": 9961,
      "benchmarks gpt4": 3444,
      "common methods": 5009,
      "methods effective": 20022,
      "demonstrating significant": 7587,
      "advanced llm": 1429,
      "response research": 28481,
      "research introduce": 28328,
      "approach identify": 2294,
      "media posts": 19761,
      "source large": 30564,
      "power natural": 25326,
      "focuses developing": 11667,
      "understand users": 34203,
      "benefit language": 3481,
      "assist people": 2633,
      "learnable parameters": 17517,
      "networks despite": 22709,
      "improvement achieving": 14327,
      "greatly reduces": 13276,
      "especially dealing": 9732,
      "longer context": 19195,
      "normalization parameters": 23018,
      "facilitates efficient": 10850,
      "prompts large": 26426,
      "prompt study": 26346,
      "parameters ranging": 24284,
      "computation time": 5451,
      "models compared": 21072,
      "optimization employing": 23625,
      "method enhancing": 19915,
      "additionally findings": 1285,
      "predominantly focused": 25463,
      "challenges large": 4354,
      "outdated knowledge": 23738,
      "single multihop": 30213,
      "sparql queries": 30608,
      "available evaluate": 2974,
      "direct prompting": 8216,
      "need new": 22640,
      "complex relationships": 5291,
      "fail represent": 10907,
      "reveal limitations": 28801,
      "longer narratives": 19198,
      "contexts language": 5942,
      "nlp recently": 22944,
      "skills model": 30313,
      "model development": 20468,
      "applications education": 2151,
      "questionanswering benchmark": 27082,
      "benchmark consisting": 3364,
      "helps measure": 13526,
      "freeform generation": 11913,
      "knowledge finetuning": 15852,
      "dialogue datasets": 8014,
      "math datasets": 19668,
      "build opensource": 3917,
      "efficient autonomous": 9028,
      "autonomous agent": 2929,
      "aim improve": 1645,
      "improve reasoning": 14292,
      "methods design": 20020,
      "strategy llms": 31127,
      "small llm": 30352,
      "make decisions": 19466,
      "integrate llm": 15321,
      "dataset finetune": 6991,
      "finetune base": 11281,
      "tuning llama7b": 33994,
      "indomain outdomain": 14725,
      "datasets code": 7074,
      "lifelong learning": 17823,
      "enabling efficient": 9320,
      "llms contrast": 18491,
      "contrast conventional": 6011,
      "conventional approaches": 6072,
      "effective strategy": 8897,
      "data settings": 6861,
      "newly curated": 22869,
      "curated dataset": 6469,
      "set evaluation": 29683,
      "experimental evaluation": 10384,
      "models greater": 21341,
      "parameters iii": 24256,
      "baselines tasks": 3271,
      "models dont": 21170,
      "dont learn": 8665,
      "language current": 16055,
      "gap human": 12090,
      "llms support": 18984,
      "deliberate reasoning": 7397,
      "reasoning chainofthought": 27391,
      "family llama": 10980,
      "13b llama": 126,
      "random chance": 27175,
      "baseline accuracy": 3240,
      "knowledge acquired": 15809,
      "reasoning poses": 27436,
      "llms make": 18791,
      "make task": 19485,
      "task practical": 32177,
      "problem solver": 26014,
      "use tools": 34575,
      "domains evaluate": 8618,
      "absolute accuracy": 761,
      "typically prompted": 34079,
      "prompted follow": 26368,
      "inference work": 14824,
      "work analyze": 35668,
      "benchmark comprehensive": 3359,
      "times average": 33157,
      "require multiple": 28222,
      "better tasks": 3630,
      "information flow": 14869,
      "network large": 22694,
      "icl capabilities": 13975,
      "finetuning remains": 11507,
      "enhance adaptability": 9504,
      "proves effective": 26680,
      "effective finetuning": 8870,
      "demands computing": 7414,
      "issue introducing": 15655,
      "peft approach": 24435,
      "label words": 15953,
      "anchors information": 2008,
      "network gnn": 22691,
      "experiments text": 10492,
      "tasks gpt2": 32350,
      "gpt2 llama2": 12913,
      "methods fewshot": 20037,
      "parameters compare": 24231,
      "prefix tuning": 25480,
      "efficiency analysis": 8996,
      "reliability challenges": 27948,
      "challenges hallucination": 4349,
      "studies reveal": 31281,
      "reveal highly": 28799,
      "capable llms": 4111,
      "responses query": 28508,
      "llms effectively": 18552,
      "responses propose": 28506,
      "method named": 19949,
      "unlike previous": 34397,
      "methods assess": 20007,
      "pair reference": 23970,
      "outperform strong": 23788,
      "finetuning demonstrate": 11387,
      "token consumption": 33187,
      "instructiontuned llama7b": 15294,
      "phi2 27b": 24930,
      "fewer training": 11094,
      "perform endtoend": 24485,
      "100 languages": 36,
      "factchecking tasks": 10860,
      "gpt4 gpt35turbo": 13083,
      "multilingual corpora": 22301,
      "languages compared": 16866,
      "approach mitigate": 2314,
      "languages languages": 16883,
      "continue training": 5988,
      "solely relying": 30464,
      "relying translation": 27981,
      "original capabilities": 23699,
      "limit performance": 17908,
      "crosslingual knowledge": 6414,
      "effectively improve": 8920,
      "performance leveraging": 24657,
      "source languages": 30563,
      "languages various": 16921,
      "comprehension generation": 5342,
      "enhance multilingual": 9522,
      "minimizing impact": 20200,
      "original performance": 23716,
      "performance resourcerich": 24739,
      "using lora": 34828,
      "improve task": 14300,
      "model learned": 20608,
      "efficient lowrank": 9047,
      "using lowrank": 34830,
      "based mistral7b": 3195,
      "different target": 8146,
      "domains finetuning": 8620,
      "data approach": 6601,
      "approach domain": 2263,
      "domain generalization": 8566,
      "models natural": 21744,
      "text large": 32903,
      "scale nli": 29145,
      "datasets today": 7181,
      "models improved": 21374,
      "generating synthetic": 12444,
      "nli data": 22912,
      "tokens labels": 33235,
      "accuracy models": 899,
      "new downstream": 22796,
      "data improves": 6733,
      "improves average": 14370,
      "average compared": 3013,
      "compared training": 5178,
      "training best": 33444,
      "t5 xxl": 31967,
      "personal experiences": 24881,
      "thinking allows": 33068,
      "focuses aspects": 11666,
      "information ii": 14872,
      "scenarios test": 29221,
      "results scaling": 28674,
      "scaling lms": 29173,
      "performance boosts": 24536,
      "scenarios ii": 29208,
      "recalling relevant": 27471,
      "finding needle": 11225,
      "needle haystack": 22650,
      "usage large": 34505,
      "efficient inference": 9039,
      "inference models": 14794,
      "methods limited": 20062,
      "ability scale": 720,
      "scale larger": 29140,
      "adapt different": 1163,
      "different hyperparameters": 8081,
      "dynamic programming": 8762,
      "tokens achieve": 33216,
      "novel sampling": 23105,
      "different decoding": 8068,
      "lms proven": 19106,
      "proven powerful": 26675,
      "powerful tools": 25354,
      "research model": 28337,
      "strands research": 31096,
      "tasks benchmark": 32251,
      "pythia models": 26901,
      "linear probing": 17991,
      "use study": 34573,
      "study learning": 31356,
      "negative polarity": 22662,
      "semeval2024 task": 29492,
      "translation paper": 33842,
      "asian languages": 2550,
      "task build": 32088,
      "model identify": 20567,
      "sentences target": 29559,
      "models extensively": 21249,
      "used machine": 34610,
      "using combination": 34753,
      "par baseline": 24150,
      "1st place": 205,
      "2nd place": 300,
      "massive multitask": 19627,
      "focus language": 11650,
      "knowledgeintensive tasks": 15932,
      "texts evaluating": 32991,
      "challenging limited": 4386,
      "limited availability": 17940,
      "models notably": 21758,
      "bloomz mt0": 3797,
      "struggle achieve": 31237,
      "score 50": 29265,
      "achieves score": 1061,
      "promising paradigm": 26291,
      "boost model": 3816,
      "efficiency large": 9005,
      "efforts explored": 9090,
      "high sparsity": 13584,
      "performance specifically": 24764,
      "activation distribution": 1139,
      "llama213b respectively": 18199,
      "llm parallel": 18345,
      "achieving superior": 1109,
      "length sequences": 17711,
      "conducted empirical": 5632,
      "computational budgets": 5455,
      "stochastic beam": 31069,
      "beam search": 3300,
      "reducing computational": 27744,
      "cost llm": 6247,
      "llama opt": 18136,
      "models showing": 21968,
      "methods consistently": 20012,
      "potential various": 25307,
      "data vital": 6909,
      "key enhancing": 15764,
      "datasets tend": 7180,
      "bilingual english": 3708,
      "corpus contains": 6177,
      "results llama": 28639,
      "llama baichuan": 18077,
      "especially zeroshot": 9751,
      "opensource resource": 23542,
      "models external": 21250,
      "concerns raised": 5545,
      "regarding behavior": 27809,
      "llms ways": 19043,
      "investigate llm": 15586,
      "likert scale": 17904,
      "scale evaluate": 29133,
      "questions using": 27134,
      "finetuned llama27b": 11336,
      "ask generate": 2552,
      "different roles": 8134,
      "types llms": 34067,
      "humans work": 13933,
      "choice questions": 4555,
      "questions llms": 27122,
      "cognitive abilities": 4874,
      "abilities knowledge": 627,
      "size paper": 30269,
      "simulated conversations": 30186,
      "opensourced llms": 23554,
      "llms families": 18603,
      "instructing llms": 15140,
      "stability models": 30847,
      "settings different": 29731,
      "personas llms": 24893,
      "conversation length": 6091,
      "need future": 22631,
      "paper provides": 24126,
      "augmentation da": 2797,
      "field study": 11153,
      "study novel": 31369,
      "novel techniques": 23118,
      "artificial data": 2529,
      "demonstrated great": 7523,
      "small data": 30337,
      "paper challenge": 24019,
      "performing better": 24834,
      "time finetuning": 33126,
      "significant contribution": 29974,
      "performs best": 24841,
      "generate data": 12269,
      "data close": 6633,
      "generation conversational": 12481,
      "agents chatgpt": 1568,
      "does work": 8542,
      "work classical": 35676,
      "model embeddings": 20485,
      "recent approaches": 27510,
      "approaches improving": 2376,
      "improving extraction": 14410,
      "largely focused": 17308,
      "data backbone": 6620,
      "backbone pretrained": 3056,
      "contain information": 5828,
      "information tokens": 14919,
      "tokens appear": 33218,
      "appear later": 2113,
      "input address": 15003,
      "address limitation": 1340,
      "simple approach": 30141,
      "extract embeddings": 10744,
      "tokens encode": 33222,
      "encode information": 9337,
      "later tokens": 17415,
      "tokens allowing": 33217,
      "leverage highquality": 17749,
      "highquality llms": 13696,
      "mistral7b model": 20239,
      "compared prior": 5165,
      "models leverage": 21444,
      "involves understanding": 15633,
      "core contributions": 6150,
      "multiple experts": 22392,
      "developments generative": 7979,
      "study utility": 31408,
      "setting paper": 29726,
      "study popular": 31373,
      "gpt35 llama2": 13028,
      "llama2 palm2": 18191,
      "finally perform": 11199,
      "llms summarize": 18981,
      "summarize findings": 31629,
      "training memoryefficient": 33562,
      "proposed address": 26588,
      "potentially explaining": 25312,
      "exhibits significant": 10250,
      "finetuning various": 11554,
      "inspired success": 15101,
      "zerothorder optimization": 36001,
      "optimization approach": 23623,
      "approach applies": 2237,
      "carefully chosen": 4172,
      "subset parameters": 31451,
      "parameters propose": 24281,
      "additionally develop": 1280,
      "a100 gpu": 616,
      "achieves absolute": 1027,
      "decoding method": 7276,
      "tokens large": 33237,
      "widespread application": 35595,
      "process address": 26049,
      "light limitations": 17828,
      "chimera novel": 4538,
      "framework specifically": 11894,
      "introduce lightweight": 15513,
      "previously generated": 25898,
      "ensure accuracy": 9601,
      "demonstrates impressive": 7560,
      "results achieving": 28567,
      "achieving average": 1084,
      "compared vanilla": 5181,
      "framework significantly": 11893,
      "improving efficiency": 14409,
      "demonstrating remarkable": 7586,
      "tasks various": 32547,
      "exploration knowledge": 10564,
      "seeks evaluate": 29361,
      "achieve conduct": 957,
      "evaluation prominent": 9995,
      "zephyr models": 35932,
      "require fewer": 28216,
      "fewer resources": 11091,
      "making suitable": 19516,
      "models best": 21022,
      "tasks domain": 32301,
      "levels comparable": 17739,
      "indicates pretraining": 14700,
      "pretraining extensive": 25797,
      "models findings": 21273,
      "underscore potential": 34176,
      "valuable resource": 35016,
      "resource understanding": 28417,
      "various aspects": 35075,
      "lack large": 15996,
      "probability target": 25973,
      "vicuna mistral": 35253,
      "comprehension capability": 5341,
      "incurs substantial": 14667,
      "lead potential": 17468,
      "llms robust": 18927,
      "expensive pretraining": 10365,
      "llms target": 18994,
      "lightweight continual": 17838,
      "comprises main": 5432,
      "main stages": 19400,
      "finetuning target": 11542,
      "parameter space": 24200,
      "matrices finetuning": 19690,
      "prominent chat": 26265,
      "architectures scales": 2471,
      "domains demonstrate": 8617,
      "demonstrate superiority": 7505,
      "models heavily": 21349,
      "highquality pretraining": 13697,
      "data order": 6789,
      "improve data": 14260,
      "data quality": 6823,
      "manually curate": 19567,
      "propose data": 26502,
      "probing evaluation": 25981,
      "data proposed": 6814,
      "use framework": 34531,
      "example use": 10112,
      "improving data": 14405,
      "quality automated": 26943,
      "pretraining gpt2": 25799,
      "benchmark evaluate": 3383,
      "evaluate large": 9841,
      "intellectual property": 15348,
      "limited understanding": 17972,
      "property ip": 26481,
      "domain paper": 8581,
      "supervised finetuned": 31673,
      "data evaluate": 6696,
      "benchmark experimental": 3386,
      "noticeable margin": 23040,
      "compared chatgpt": 5125,
      "specially curated": 30680,
      "challenging problem": 4393,
      "llms process": 18870,
      "texts paper": 32998,
      "detection method": 7870,
      "experiments representative": 10476,
      "subset neurons": 31450,
      "furthermore showcase": 12012,
      "output language": 23868,
      "language llms": 16112,
      "important evidence": 14200,
      "generalist models": 12198,
      "demonstrated capabilities": 7516,
      "llms plain": 18852,
      "data remains": 6835,
      "limited investigation": 17951,
      "investigation reveals": 15615,
      "ability process": 715,
      "model average": 20384,
      "developed comprehensive": 7923,
      "comprehensive instruction": 5381,
      "comprising 11": 5435,
      "11 million": 73,
      "train series": 33373,
      "generalization novel": 12224,
      "alignment pretrained": 1777,
      "text originating": 32913,
      "automatically construct": 2906,
      "containing 20k": 5834,
      "llama2 despite": 18172,
      "use recent": 34565,
      "recent knowledge": 27522,
      "investigate various": 15600,
      "alignment experiments": 1759,
      "aligning models": 1750,
      "finetuning models": 11455,
      "representations using": 28178,
      "attention mask": 2724,
      "training transformerbased": 33641,
      "taskspecific soft": 32569,
      "models attention": 20993,
      "easy implement": 8801,
      "text reasonable": 32928,
      "word level": 35641,
      "evaluate lms": 9847,
      "lms ability": 19067,
      "english speakers": 9491,
      "task seen": 32191,
      "gpt2 bloom": 12876,
      "bloom chatgpt": 3784,
      "calibration error": 3977,
      "variational learning": 35047,
      "learning effective": 17562,
      "effective large": 8880,
      "empirical evidence": 9224,
      "optimizer called": 23650,
      "networks gpt2": 22712,
      "generalization error": 12213,
      "evidence support": 10064,
      "model representations": 20753,
      "methods successfully": 20099,
      "disentangle roles": 8306,
      "models dataset": 21123,
      "use resulting": 34566,
      "conceptual framework": 5534,
      "define new": 7363,
      "distributed representations": 8387,
      "release benchmark": 27900,
      "layer dropping": 17424,
      "training reducing": 33597,
      "approach designed": 2257,
      "designed reduce": 7743,
      "efficiency training": 9018,
      "specifically utilizing": 30761,
      "loss level": 19247,
      "new generation": 22807,
      "model series": 20773,
      "instructiontuned versions": 15300,
      "thorough evaluations": 33073,
      "multiturn dialogues": 22464,
      "open model": 23408,
      "parameters significant": 24288,
      "metrics compared": 20133,
      "based framework": 3163,
      "inserting new": 15059,
      "models explore": 21240,
      "explore approach": 10573,
      "steps propose": 31060,
      "propose leverage": 26523,
      "plms bert": 25041,
      "flant5 llama": 11597,
      "datasets created": 7087,
      "created using": 6362,
      "entity linking": 9645,
      "best settings": 3579,
      "framework use": 11901,
      "use finetuned": 34530,
      "finetuned plm": 11350,
      "task propose": 32182,
      "shows advantages": 29921,
      "encouraging performance": 9403,
      "follow complex": 11676,
      "complex domainspecific": 5274,
      "agents despite": 1569,
      "despite llms": 7794,
      "range realworld": 27209,
      "opensource llama": 23515,
      "gemini llms": 12140,
      "varies different": 35053,
      "insights suggest": 15083,
      "suggest need": 31576,
      "human automated": 13796,
      "major bottleneck": 19437,
      "time large": 33130,
      "present collection": 25519,
      "model sets": 20777,
      "point improvement": 25065,
      "language resources": 16819,
      "include new": 14447,
      "open license": 23405,
      "including research": 14516,
      "commercial usage": 4997,
      "contribute advancement": 6028,
      "language technology": 16833,
      "focuses task": 11669,
      "generated response": 12384,
      "specific query": 30711,
      "task new": 32167,
      "new llm": 22818,
      "new query": 22836,
      "significantly increase": 30064,
      "impractical realworld": 14227,
      "problem paper": 26000,
      "observe llms": 23231,
      "gpt4 finetuning": 13077,
      "generate response": 12319,
      "required output": 28231,
      "limited certain": 17943,
      "llms improving": 18687,
      "solution problem": 30476,
      "used augment": 34585,
      "automatically generating": 2916,
      "problem provide": 26008,
      "problem inspired": 25992,
      "socratic questioning": 30445,
      "specific ways": 30725,
      "effectively avoid": 8912,
      "prompting methods": 26389,
      "network mechanisms": 22697,
      "nodes edges": 22983,
      "contrast existing": 6012,
      "activation patching": 1141,
      "allows efficiently": 1813,
      "single forward": 30201,
      "carefully design": 4177,
      "extract information": 10745,
      "specific types": 30722,
      "experiment llama": 10378,
      "role attention": 28953,
      "llms text": 19002,
      "llms known": 18738,
      "parameter counts": 24177,
      "limited memory": 17956,
      "proposed solution": 26619,
      "target models": 32056,
      "framework train": 11897,
      "model llama": 20618,
      "distillation additional": 8335,
      "finetuning step": 11537,
      "step use": 31051,
      "instructionresponse pairs": 15240,
      "gradient method": 13189,
      "method reinforcement": 19966,
      "24times speedup": 269,
      "tasks taskspecific": 32523,
      "requires identifying": 28255,
      "information mitigate": 14888,
      "systems propose": 31915,
      "compare multiple": 5110,
      "llms existing": 18583,
      "existing event": 10274,
      "demonstrates strong": 7572,
      "fewshot llms": 11115,
      "llms iteratively": 18731,
      "directly prompting": 8242,
      "inherent limitations": 14950,
      "operational efficiency": 23567,
      "design lack": 7706,
      "lack domainspecific": 15986,
      "adapt llms": 1166,
      "llms solving": 18960,
      "manner specifically": 19550,
      "establish comprehensive": 9768,
      "twophase learning": 34042,
      "learning strategy": 17657,
      "convergence behavior": 6084,
      "behavior model": 3320,
      "350m model": 344,
      "gpt4 turbo": 13125,
      "sets finetuned": 29719,
      "instructions available": 15243,
      "feedback reinforcement": 11069,
      "learning automatically": 17542,
      "systems online": 31910,
      "online learning": 23365,
      "potential improve": 25263,
      "feedback generation": 11062,
      "requires models": 28259,
      "understand problem": 34200,
      "effectively use": 8933,
      "humanwritten llmgenerated": 13937,
      "feedback second": 11073,
      "alignment using": 1783,
      "augmented dataset": 2812,
      "llama opensource": 18135,
      "areas future": 2485,
      "evolving field": 10081,
      "linguistic descriptions": 18010,
      "mathematical formulation": 19680,
      "presents formidable": 25585,
      "formidable challenge": 11764,
      "study compares": 31307,
      "oneshot settings": 23352,
      "settings task": 29744,
      "performance particularly": 24713,
      "central research": 4268,
      "specialized datasets": 30669,
      "notable gap": 23025,
      "capabilities smaller": 4070,
      "llama27b compared": 18205,
      "larger counterparts": 17319,
      "lengthy complex": 17717,
      "research achieving": 28286,
      "achieving f1score": 1091,
      "solely based": 30462,
      "problem description": 25989,
      "benchmark current": 3369,
      "current capabilities": 6486,
      "llms novel": 18820,
      "novel application": 23055,
      "application area": 2127,
      "groundwork future": 13296,
      "future improvements": 12033,
      "llms reflect": 18899,
      "lexical semantics": 17802,
      "architectures paper": 2469,
      "specifically investigate": 30747,
      "llm llama2": 18335,
      "contextualized word": 5963,
      "identification task": 13997,
      "lower layers": 19288,
      "contrast models": 6014,
      "increase performance": 14603,
      "unique model": 34361,
      "model design": 20461,
      "language abilities": 16033,
      "successfully improve": 31541,
      "generation performance": 12570,
      "retaining original": 28719,
      "adaptation large": 1180,
      "model foundation": 20536,
      "model vs": 20862,
      "instruction model": 15171,
      "model providing": 20737,
      "analysis present": 1948,
      "resources publicly": 28444,
      "play crucial": 25017,
      "information various": 14925,
      "types user": 34074,
      "final answer": 11176,
      "construct instruction": 5798,
      "diverse tabular": 8465,
      "generate accurate": 12261,
      "work underscores": 35796,
      "underscores importance": 34180,
      "abilities model": 634,
      "release dataset": 27906,
      "layers llms": 17441,
      "llms necessary": 18813,
      "inference phase": 14797,
      "llms expensive": 18585,
      "llms utilize": 19031,
      "generalization incontext": 12216,
      "paper try": 24143,
      "try answer": 33946,
      "shallow layers": 29774,
      "deep layers": 7322,
      "layers tasks": 17446,
      "experiments wellknown": 10502,
      "tasks maintaining": 32412,
      "maintaining comparable": 19419,
      "performance additionally": 24518,
      "additionally method": 1291,
      "comparable superior": 5092,
      "progress llms": 26216,
      "drawing inspiration": 8727,
      "llms capabilities": 18456,
      "significantly affect": 30033,
      "affect llms": 1537,
      "tasks particular": 32446,
      "gpt35 llama": 13027,
      "exhibit higher": 10220,
      "models adopt": 20965,
      "specific personas": 30709,
      "finetuned curated": 11305,
      "incur significant": 14659,
      "hindering widespread": 13718,
      "data contributes": 6660,
      "rtx 4090": 29002,
      "a100 40gb": 614,
      "single rtx": 30221,
      "highquality instructions": 13694,
      "tasks resulting": 32488,
      "designed address": 7719,
      "report provides": 28124,
      "provides overview": 26761,
      "additional pretraining": 1263,
      "word problem": 35644,
      "ambiguous contexts": 1875,
      "presents new": 25587,
      "evaluating llm": 9904,
      "develop dataset": 7913,
      "questions categories": 27100,
      "developed evaluation": 7925,
      "evaluation methodology": 9972,
      "text similarity": 32940,
      "mathematical expression": 19679,
      "llama claude": 18085,
      "avoid hallucination": 3038,
      "hallucination code": 13373,
      "models minimal": 21723,
      "minimal human": 20186,
      "serving large": 29664,
      "issue particularly": 15658,
      "particularly pronounced": 24354,
      "language like": 16110,
      "directly translating": 8245,
      "selfinstruct method": 29425,
      "based gpt4": 3170,
      "gpt4 translate": 13124,
      "quality gpt4": 26966,
      "construct evaluation": 5797,
      "benchmark containing": 3365,
      "80 questions": 554,
      "automatically assess": 2904,
      "gpt4 selfinstruct": 13115,
      "selfinstruct data": 29424,
      "base pretrained": 3129,
      "gpt35 davinci003": 13020,
      "benchmark released": 3408,
      "investigate basic": 15577,
      "standard benchmark": 30869,
      "typologically diverse": 34085,
      "models respond": 21937,
      "answering accuracy": 2056,
      "accuracy llms": 896,
      "use models": 34552,
      "differences models": 8039,
      "languages models": 16899,
      "explore differences": 10580,
      "identifying possible": 14022,
      "highlevel semantic": 13625,
      "semantic concepts": 29451,
      "prediction objective": 25430,
      "bias gradient": 3646,
      "linear representation": 17995,
      "using llama2": 34821,
      "simplified model": 30178,
      "cuttingedge large": 6571,
      "models employed": 21194,
      "specifically compare": 30730,
      "divideandconquer approach": 8488,
      "thought cot": 33079,
      "certain models": 4277,
      "gpt4 claude21": 13065,
      "accuracy rates": 905,
      "detrimental effects": 7905,
      "struggle accurately": 31236,
      "despite showing": 7811,
      "factors impact": 10870,
      "tasks extensive": 32325,
      "task complexity": 32097,
      "information density": 14858,
      "tasks potential": 32451,
      "future large": 12035,
      "underscores potential": 34182,
      "potential revolutionize": 25294,
      "efficiency deployment": 9000,
      "sophisticated models": 30526,
      "size computational": 30241,
      "models surprisingly": 22034,
      "compact powerful": 5067,
      "paper conducts": 24025,
      "conducts comprehensive": 5647,
      "intrinsic understanding": 15494,
      "potential limitations": 25272,
      "obtain significant": 23252,
      "decoderonly pretrained": 7264,
      "proved effective": 26669,
      "research question": 28354,
      "performances existing": 24820,
      "truly understand": 33929,
      "sentiment understanding": 29572,
      "scenarios especially": 29203,
      "llms shows": 18946,
      "tasks chat": 32261,
      "proposes new": 26629,
      "task aiming": 32075,
      "ability understanding": 725,
      "scenarios particularly": 29214,
      "task introduces": 32145,
      "auxiliary task": 2961,
      "factual hallucination": 10883,
      "hallucination problem": 13381,
      "hallucination issues": 13377,
      "error correction": 9711,
      "llama 2based": 18062,
      "synthetic errors": 31856,
      "akin human": 1678,
      "correction models": 6199,
      "models gains": 21302,
      "gpt4 results": 13114,
      "results synthetic": 28695,
      "llm queries": 18356,
      "help users": 13515,
      "product reviews": 26169,
      "inference highly": 14780,
      "queries present": 27024,
      "keyvalue kv": 15794,
      "kv cache": 15948,
      "inference engine": 14774,
      "model serving": 20775,
      "endtoend latency": 9435,
      "sql queries": 30840,
      "terms effectiveness": 32744,
      "llms transformerbased": 19008,
      "extraordinary capabilities": 10786,
      "paper test": 24142,
      "prompts bring": 26406,
      "potential task": 25301,
      "dataset revealed": 7035,
      "interesting observations": 15409,
      "task second": 32190,
      "provide feedback": 26701,
      "enhance quality": 9527,
      "languages despite": 16870,
      "despite considerable": 7772,
      "considerable advancements": 5704,
      "aims bridge": 1659,
      "llms covering": 18496,
      "containing total": 5837,
      "quality quantity": 26974,
      "data synthetic": 6887,
      "opensource pipeline": 23539,
      "diverse sources": 8464,
      "mixtral models": 20275,
      "models create": 21114,
      "toxicity alignment": 33316,
      "prompts multiple": 26433,
      "generate nontoxic": 12308,
      "llms establish": 18568,
      "languages data": 16868,
      "work released": 35772,
      "development reliable": 7972,
      "family caregivers": 10974,
      "urgent need": 34496,
      "quality care": 26945,
      "models potentially": 21827,
      "potentially used": 25316,
      "educational tools": 8846,
      "care study": 4166,
      "study aimed": 31297,
      "aimed develop": 1651,
      "requires fewer": 28252,
      "resources evaluate": 28435,
      "compared large": 5142,
      "rag framework": 27158,
      "improving quality": 14420,
      "falcon 7b": 10939,
      "parameters larger": 24265,
      "benchmark developed": 3381,
      "caregivers individuals": 4182,
      "using benchmark": 34739,
      "evaluating language": 9897,
      "provide accurate": 26682,
      "study measures": 31360,
      "memory access": 19801,
      "access language": 818,
      "tasks particularly": 32447,
      "underlying knowledge": 34154,
      "remain elusive": 27983,
      "lm gpt2": 19057,
      "synthetic tasks": 31860,
      "memorized content": 19796,
      "realistic scenarios": 27316,
      "answering code": 2057,
      "reproduce experiments": 28199,
      "values using": 35027,
      "fixed vocabulary": 11580,
      "existing transformerbased": 10322,
      "family ranging": 10984,
      "parameters large": 24261,
      "datasets comprising": 7080,
      "local models": 19134,
      "methods datasets": 20018,
      "datasets relative": 7168,
      "greatly simplify": 13278,
      "design generative": 7704,
      "iot devices": 15638,
      "llms stand": 18967,
      "era artificial": 9693,
      "directly deploying": 8236,
      "llms resourceconstrained": 18916,
      "resourceconstrained hardware": 28420,
      "cost paper": 6252,
      "models key": 21409,
      "key design": 15760,
      "transformer decoders": 33714,
      "given computational": 12743,
      "solving mathematical": 30513,
      "models termed": 22052,
      "autoregressive transformer": 2956,
      "compared 350m": 5118,
      "350m parameter": 345,
      "parameter opt": 24194,
      "nvidia jetson": 23195,
      "available soon": 3001,
      "information extracted": 14866,
      "previous state": 25879,
      "baseline results": 3258,
      "highlight critical": 13627,
      "critical role": 6393,
      "systems performance": 31913,
      "processing interpreting": 26106,
      "suggest promising": 31578,
      "task datasets": 32103,
      "family lightweight": 10979,
      "stateofthe art": 30918,
      "gemma models": 12146,
      "sizes models": 30300,
      "parameters provide": 24283,
      "models 11": 20919,
      "models alongside": 20977,
      "detailed description": 7836,
      "improving safety": 14422,
      "frontier models": 11933,
      "innovations language": 14993,
      "large curated": 16939,
      "play vital": 25020,
      "relatively little": 27888,
      "corpora paper": 6169,
      "paper compare": 24021,
      "relevant large": 27944,
      "intrinsic evaluation": 15490,
      "taken different": 32024,
      "different corpora": 8062,
      "practical impact": 25365,
      "differences training": 8041,
      "training specific": 33619,
      "clear differences": 4647,
      "training lms": 33555,
      "teaching llms": 32593,
      "instructions reinforcement": 15269,
      "learning development": 17558,
      "challenges stemming": 4376,
      "reliance human": 27961,
      "rlhf framework": 28907,
      "paradigm work": 24160,
      "llms following": 18619,
      "directly generate": 8238,
      "generation highquality": 12516,
      "excessive reliance": 10180,
      "reliance external": 27960,
      "advanced models": 1432,
      "way single": 35448,
      "rlhf stages": 28910,
      "highlight key": 13632,
      "key advantages": 15753,
      "instructions compared": 15246,
      "compared strong": 5173,
      "improved model": 14315,
      "privacy protection": 25953,
      "investigating performance": 15611,
      "opened new": 23454,
      "development new": 7963,
      "new types": 22864,
      "techniques used": 32670,
      "systems study": 31920,
      "using rouge": 34906,
      "rouge bleu": 28978,
      "different datasets": 8067,
      "models rag": 21884,
      "decrease performance": 7298,
      "based cosine": 3146,
      "generation summarization": 12608,
      "test hypothesis": 32770,
      "hypothesis conducted": 13964,
      "effective robust": 8894,
      "performing task": 24839,
      "important step": 14212,
      "dataset evaluated": 6983,
      "assess robustness": 2604,
      "scenarios results": 29219,
      "achieved highest": 1009,
      "robustness compared": 28944,
      "weather conditions": 35470,
      "llama achieved": 18070,
      "achieved good": 1005,
      "good results": 12824,
      "certain conditions": 4273,
      "llms neural": 18816,
      "performed large": 24830,
      "tests llm": 32808,
      "mathematical abilities": 19678,
      "domain finally": 8563,
      "human studies": 13865,
      "models case": 21046,
      "resources numerous": 28441,
      "llms mllms": 18802,
      "languages lrls": 16895,
      "lrls study": 19322,
      "strategies enhance": 31104,
      "bilingual data": 3707,
      "quantitatively evaluated": 26998,
      "llms tasks": 18996,
      "based human": 3173,
      "evaluation gpt4": 9958,
      "results showed": 28678,
      "qualitative analyses": 26930,
      "previously proposed": 25900,
      "monolingual models": 22209,
      "represent diverse": 28133,
      "probing task": 25983,
      "models opt": 21774,
      "models equally": 21210,
      "augmented finetuning": 2814,
      "finetuning scaling": 11516,
      "faces significant": 10830,
      "memory constraints": 19808,
      "multiple gpus": 22394,
      "efficient parameter": 9053,
      "resource management": 28414,
      "limited gpu": 17949,
      "gpu resources": 13177,
      "resources experiments": 28436,
      "runtime compared": 29017,
      "vram gpu": 35399,
      "probing classifiers": 25980,
      "applications automated": 2141,
      "model instead": 20585,
      "propose directly": 26504,
      "extraction capabilities": 10764,
      "efficient simultaneous": 9057,
      "introduce approach": 15498,
      "finetuning incurring": 11420,
      "minimal additional": 20182,
      "additional computational": 1252,
      "maintains high": 19432,
      "token generation": 33191,
      "baseline using": 3260,
      "using separate": 34908,
      "ner model": 22677,
      "challenge dataset": 4309,
      "fundamental cognitive": 11975,
      "deeply rooted": 7352,
      "llms release": 18901,
      "instances containing": 15112,
      "carefully selected": 4179,
      "manually annotated": 19563,
      "different levels": 8092,
      "dataset freely": 6994,
      "language different": 16063,
      "research suggests": 28364,
      "paper establish": 24038,
      "establish benchmark": 9767,
      "results popular": 28656,
      "task far": 32123,
      "models facto": 21257,
      "produce accurate": 26138,
      "realworld data": 27336,
      "data tends": 6890,
      "perspective llm": 24896,
      "curation pipeline": 6475,
      "automatically identifying": 2921,
      "existing data": 10266,
      "curation techniques": 6477,
      "comprehensive framework": 5379,
      "dataset trained": 7047,
      "assume access": 2659,
      "finetuning gpt35": 11414,
      "capabilities llm": 4038,
      "llm experiments": 18300,
      "reveal clear": 28791,
      "advancement generative": 1451,
      "intelligence genai": 15357,
      "diverse sectors": 8458,
      "performance computing": 24557,
      "concept generation": 5527,
      "guide autoregressive": 13343,
      "user prompts": 34666,
      "realworld evaluations": 27340,
      "llama2 llm": 18183,
      "data research": 6839,
      "critical step": 6394,
      "step aligning": 31037,
      "potential mitigating": 25278,
      "expanding domain": 10341,
      "distillation efficient": 8337,
      "prompt compression": 26315,
      "compress prompts": 5403,
      "information entropy": 14864,
      "challenge information": 4315,
      "fail capture": 10900,
      "essential information": 9759,
      "information needed": 14890,
      "llm compress": 18282,
      "crucial information": 6442,
      "use transformer": 34578,
      "bidirectional context": 3691,
      "lower latency": 19287,
      "explicitly learning": 10548,
      "outofdomain datasets": 23752,
      "datasets including": 7132,
      "ability different": 661,
      "llms additionally": 18425,
      "additionally model": 1293,
      "faster existing": 10999,
      "methods accelerating": 19999,
      "compression ratios": 5426,
      "tasks requires": 32484,
      "methods different": 20021,
      "users flexibly": 34690,
      "100 llms": 37,
      "effectiveness framework": 8947,
      "instructionfinetuned large": 15218,
      "analyze models": 1998,
      "reasoning context": 27396,
      "political science": 25095,
      "aware instruction": 3042,
      "translation capabilities": 33821,
      "commercial translation": 4996,
      "translation systems": 33853,
      "models mitigate": 21728,
      "llms translation": 19010,
      "instructions especially": 15250,
      "design twostage": 7716,
      "ability especially": 663,
      "llms maximum": 18797,
      "dataset elicit": 6979,
      "capabilities second": 4068,
      "samples randomly": 29087,
      "translation directions": 33825,
      "samples experiments": 29076,
      "benchmarks llama": 3458,
      "zeroshot directions": 35967,
      "compared competitive": 5126,
      "competitive baseline": 5220,
      "llama method": 18125,
      "translation quality": 33847,
      "agent trajectories": 1564,
      "abilities reasoning": 638,
      "recently efforts": 27592,
      "train language": 33363,
      "agents performance": 1573,
      "diverse prompting": 8450,
      "framework enables": 11845,
      "central role": 4269,
      "multiple rounds": 22416,
      "language agent": 16040,
      "using qlora": 34893,
      "qlora finetuning": 26924,
      "matches human": 19652,
      "iterative refinement": 15688,
      "lead performance": 17467,
      "agents significantly": 1574,
      "existing techniques": 10318,
      "prompting gpt4": 26376,
      "fully finetuned": 11955,
      "emergence numerous": 9175,
      "numerous large": 23184,
      "generation key": 12529,
      "key task": 15786,
      "intrinsic properties": 15493,
      "properties large": 26474,
      "gpt2 dialogpt": 12883,
      "chatgpt flant5": 4469,
      "sizes small": 30304,
      "small medium": 30354,
      "medium large": 19776,
      "propose different": 26503,
      "models analysis": 20979,
      "shows improvement": 29927,
      "quality datasets": 26951,
      "generating counter": 12417,
      "counter speech": 6291,
      "speech models": 30784,
      "models metrics": 21722,
      "speech generation": 30781,
      "models exponentially": 21243,
      "encounter significant": 9394,
      "approach employs": 2268,
      "lowrank matrix": 19307,
      "accurate approximation": 920,
      "training downstream": 33503,
      "memory savings": 19828,
      "improves downstream": 14372,
      "boosting llms": 3824,
      "llms currently": 18502,
      "tasks realworld": 32473,
      "lowdata regime": 19279,
      "making finetuning": 19503,
      "strategy uses": 31130,
      "small seed": 30366,
      "seed dataset": 29353,
      "augmenting additional": 2823,
      "used finetuning": 34602,
      "student llm": 31253,
      "initial seed": 14964,
      "incorrect data": 14583,
      "examples llm": 10133,
      "enhances performance": 9551,
      "achieve improvements": 970,
      "dataset 326": 6934,
      "regular finetuning": 27826,
      "regime using": 27816,
      "using llama27b": 34823,
      "contemporary large": 5842,
      "llms engage": 18563,
      "deploy llms": 7630,
      "interaction history": 15385,
      "llm prompt": 18353,
      "using variety": 34936,
      "prompt designs": 26319,
      "models robustly": 21949,
      "gpt4 chainofthought": 13061,
      "did result": 8031,
      "result robust": 28545,
      "including chainofthought": 14463,
      "dataset curation": 6968,
      "achieve best": 945,
      "models foundation": 21293,
      "leverages llms": 17770,
      "llms annotate": 18433,
      "large unlabeled": 17284,
      "transformer encoders": 33716,
      "encoders like": 9378,
      "approach slightly": 2340,
      "scaling behaviors": 29159,
      "work create": 35685,
      "outperforming openais": 23801,
      "checkpoint publicly": 4520,
      "code facilitate": 4750,
      "use natural": 34554,
      "potential benefits": 25244,
      "models billions": 21028,
      "t5 existing": 31942,
      "models size": 21983,
      "performance sentence": 24747,
      "positive correlation": 25193,
      "particularly noteworthy": 24351,
      "model available": 20383,
      "methods rely": 20086,
      "industry applications": 14745,
      "remarkable zeroshot": 28064,
      "examples introduce": 10131,
      "comprehensive comparison": 5360,
      "shows approach": 29922,
      "flant5 11b": 11592,
      "stateoftheart joint": 30934,
      "learning gpt35": 17574,
      "gpt35 175b": 13013,
      "finetuning requires": 11510,
      "requires training": 28264,
      "data yield": 6916,
      "yield comparable": 35908,
      "accurate information": 925,
      "literature review": 18045,
      "review process": 28830,
      "retrieval using": 28758,
      "using stateoftheart": 34920,
      "structured text": 31229,
      "analysis techniques": 1971,
      "key elements": 15763,
      "analyze challenges": 1989,
      "gpt35 textdavinci003": 13035,
      "accuracy 58": 859,
      "evaluating risks": 9913,
      "ongoing dialogue": 23356,
      "major challenges": 19440,
      "llm activations": 18264,
      "including truthfulqa": 14525,
      "metric improvement": 20122,
      "improvement baseline": 14331,
      "kullbackleibler divergence": 15946,
      "hallucination detection": 13376,
      "adoption large": 1408,
      "llms facilitated": 18599,
      "significant concern": 29971,
      "rag emerged": 27157,
      "emerged highly": 9160,
      "highly promising": 13666,
      "text produced": 32921,
      "produced llms": 26157,
      "retrieved documents": 28773,
      "ukraine war": 34093,
      "aigenerated text": 1635,
      "unable accurately": 34101,
      "introduces new": 15543,
      "new type": 22863,
      "learning mtl": 17613,
      "incorporating stateoftheart": 14579,
      "compared sota": 5168,
      "rank llms": 27224,
      "hallucinations generate": 13390,
      "llms answer": 18434,
      "users information": 34692,
      "conversational context": 6097,
      "approaches model": 2385,
      "information needs": 14891,
      "query use": 27032,
      "retrieval paper": 28750,
      "retrieval methods": 28746,
      "methods leverage": 20059,
      "generating appropriate": 12410,
      "appropriate response": 2407,
      "generate multiple": 12303,
      "implement evaluate": 14158,
      "models utilizing": 22107,
      "utilizing various": 34977,
      "llama2 chat": 18166,
      "based gpt": 3166,
      "reveal effectiveness": 28795,
      "developing generative": 7944,
      "field generative": 11137,
      "intelligence gai": 15356,
      "chatbots chatgpt": 4452,
      "google bard": 12826,
      "including education": 14475,
      "potential higher": 25260,
      "acceptance model": 813,
      "method encompasses": 19911,
      "literature databases": 18041,
      "critical elements": 6387,
      "main focus": 19396,
      "given prompt": 12762,
      "prompt like": 26333,
      "additionally observed": 1295,
      "final layer": 11177,
      "correct predictions": 6193,
      "improve factual": 14266,
      "neural machine translation": 22732,
      "using pretrained language": 34879,
      "pretrained language models": 25664,
      "language models lms": 16596,
      "models lms various": 21683,
      "lms various natural": 19123,
      "various natural language": 35125,
      "natural language processing": 22535,
      "language processing tasks": 16800,
      "tasks work introduce": 32554,
      "machine translation nmt": 19361,
      "language models large": 16404,
      "models large language": 21421,
      "large language models": 16989,
      "language models range": 16677,
      "gpt2 language model": 12909,
      "masked language model": 19611,
      "masked language models": 19616,
      "language models mlms": 16627,
      "nlp tasks instead": 22955,
      "autoregressive language models": 2945,
      "language models like": 16416,
      "models like gpt2": 21453,
      "constrained text generation": 5782,
      "largescale pretrained language": 17374,
      "language models demonstrated": 16299,
      "models demonstrated impressive": 21136,
      "demonstrated impressive performance": 7529,
      "text generation task": 32883,
      "commonsense reasoning given": 5041,
      "task generate coherent": 32131,
      "stateoftheart text generation": 30999,
      "text generation models": 32878,
      "language models gpt": 16368,
      "models achieved stateoftheart": 20952,
      "achieved stateoftheart results": 1019,
      "et al 2018": 9791,
      "model improve performance": 20572,
      "performance complex problems": 24556,
      "et al 2016": 9790,
      "task model trained": 32160,
      "model trained scratch": 20836,
      "setting new stateoftheart": 29725,
      "tiny fraction parameters": 33173,
      "conduct thorough analysis": 5626,
      "previous work focused": 25892,
      "models large pretrained": 21426,
      "large pretrained language": 17260,
      "language models led": 16415,
      "results natural language": 28650,
      "natural language understanding": 22569,
      "language understanding tasks": 16850,
      "natural language models": 22533,
      "language models machine": 16614,
      "machine learning tasks": 19354,
      "models similar size": 21979,
      "parameters language model": 24259,
      "language model recently": 16196,
      "neural language models": 22727,
      "language models trained": 16732,
      "using natural language": 34850,
      "natural language queries": 22562,
      "finetuning pretrained models": 11492,
      "models answer questions": 20981,
      "code trained models": 4825,
      "data augmentation using": 6614,
      "using pretrained transformer": 34883,
      "pretrained transformer models": 25766,
      "models language model": 21416,
      "language model based": 16122,
      "model based pretrained": 20389,
      "pretrained models bert": 25719,
      "different nlp tasks": 8113,
      "nlp tasks paper": 22961,
      "transformer based pretrained": 33710,
      "pretrained models autoregressive": 25718,
      "autoencoder models bert": 2844,
      "pretrained seq2seq model": 25747,
      "generative pretrained language": 12686,
      "pretrained language model": 25657,
      "language model gpt2": 16149,
      "generative language models": 12663,
      "language models paper": 16642,
      "models paper presents": 21796,
      "paper presents empirical": 24099,
      "presents empirical study": 25583,
      "language models plms": 16653,
      "maximum likelihood estimation": 19707,
      "taskoriented dialogue systems": 32222,
      "models using data": 22102,
      "texttotext transfer transformer": 33017,
      "transfer transformer t5": 33684,
      "fewer parameters compared": 11090,
      "natural language evaluation": 22515,
      "realworld relation extraction": 27345,
      "limited training data": 17971,
      "augment training data": 2794,
      "finetuning gpt2 generate": 11412,
      "training data used": 33493,
      "11 f1 score": 71,
      "new state art": 22847,
      "f1 points average": 10816,
      "improvements nlp tasks": 14362,
      "generative language model": 12662,
      "built using gpt2": 3938,
      "transformer based models": 33709,
      "outofdomain test sets": 23755,
      "baseline future research": 3246,
      "models lms bert": 21670,
      "lms bert gpt2": 19073,
      "tasks recent work": 32475,
      "recent work focused": 27568,
      "models substantially outperform": 22023,
      "deep learning architectures": 7326,
      "paper investigate commonsense": 24069,
      "natural language inference": 22525,
      "performance language models": 24643,
      "language models finetuned": 16344,
      "multiple choice question": 22383,
      "performance experimental results": 24592,
      "powerful generative model": 25339,
      "systems paper presents": 31912,
      "paper presents fewshot": 24101,
      "language model pretraining": 16192,
      "model pretraining knowledge": 20725,
      "knowledge pretrained language": 15888,
      "language modeling tasks": 16227,
      "neural network language": 22745,
      "network language models": 22693,
      "language models lm": 16595,
      "using neural text": 34856,
      "neural text generation": 22761,
      "text generation based": 32868,
      "propose new method": 26539,
      "new method called": 22821,
      "methods significantly improve": 20095,
      "using generative language": 34786,
      "language models work": 16759,
      "learning natural language": 17620,
      "openais generative pretrained": 23439,
      "generative pretrained transformer": 12692,
      "pretrained transformer gpt2": 25759,
      "language models synthetic": 16722,
      "data used train": 6903,
      "deep learning models": 7332,
      "fields natural language": 11158,
      "language processing nlp": 16789,
      "information retrieval ir": 14909,
      "learning models like": 17611,
      "recurrent neural networks": 27684,
      "neural networks rnns": 22755,
      "long shortterm memory": 19183,
      "bidirectional encoder representations": 3693,
      "encoder representations transformers": 9355,
      "representations transformers bert": 28176,
      "deep neural network": 7338,
      "small models large": 30360,
      "recently published work": 27617,
      "work deep learning": 35687,
      "large generative language": 16951,
      "language models gpt2": 16369,
      "downstream tasks finetuning": 8693,
      "human machinegenerated text": 13846,
      "models elmo bert": 21187,
      "bert gpt gpt2": 3507,
      "models previous works": 21852,
      "largescale language models": 17360,
      "models lms able": 21669,
      "natural language generate": 22519,
      "using smaller lms": 34918,
      "guide generation large": 13346,
      "fast generation speed": 10993,
      "advanced neural language": 1438,
      "language models assessing": 16249,
      "limitations language models": 17924,
      "tradeoff language models": 33337,
      "language models including": 16386,
      "openended text generation": 23463,
      "scaling model size": 29176,
      "model size efficiently": 20789,
      "propose simple effective": 26567,
      "pretrained gpt2 model": 25650,
      "existing training data": 10321,
      "neural toxic degeneration": 22763,
      "language models pretrained": 16661,
      "pretrained neural language": 25739,
      "models lms prone": 21677,
      "controllable text generation": 6061,
      "preventing toxic degeneration": 25861,
      "gpt2 radford et": 12943,
      "radford et al": 27151,
      "et al 2019": 9792,
      "language models recently": 16690,
      "models lms demonstrated": 21671,
      "lms demonstrated impressive": 19081,
      "demonstrated impressive abilities": 7527,
      "paper propose method": 24110,
      "set linguistic features": 29693,
      "humans process language": 13931,
      "datasets compare performance": 7078,
      "bert model achieves": 3519,
      "generated language model": 12366,
      "language model like": 16162,
      "model like gpt2": 20615,
      "machine translation models": 19359,
      "code data available": 4729,
      "achieved impressive results": 1012,
      "range natural language": 27200,
      "language understanding nlu": 16844,
      "extensive experimental results": 10690,
      "experimental results method": 10401,
      "relying external knowledge": 27977,
      "outperforms baseline methods": 23807,
      "commonsense reasoning ability": 5040,
      "language models question": 16676,
      "question answering recent": 27054,
      "recent works shown": 27574,
      "shown language models": 29892,
      "generative models t5": 12680,
      "models t5 bart": 22042,
      "diverse range datasets": 8454,
      "demonstrate effectiveness methods": 7448,
      "existing pretrained models": 10310,
      "pretrained models new": 25726,
      "generated gpt2 model": 12359,
      "artificial neural networks": 2540,
      "natural language generation": 22520,
      "stateoftheart approaches demonstrate": 30922,
      "openais gpt2 model": 23443,
      "gpt2 model successfully": 12922,
      "existing work does": 10325,
      "powerful language models": 25341,
      "language models able": 16235,
      "human evaluation shows": 13808,
      "evaluation shows model": 10012,
      "model able generate": 20339,
      "make publicly available": 19480,
      "publicly available code": 26851,
      "language models bert": 16257,
      "models bert xlnet": 21021,
      "long training time": 19189,
      "extremely large batch": 10799,
      "large batch sizes": 16931,
      "computer vision tasks": 5508,
      "downstream tasks results": 8700,
      "achieves comparable performance": 1039,
      "leverage large pretrained": 17755,
      "language models perform": 16649,
      "perform downstream tasks": 24484,
      "language model parameters": 16185,
      "finetuning natural language": 11462,
      "language generation tasks": 16091,
      "trillion parameter models": 33904,
      "parameters constant computational": 24235,
      "constant computational cost": 5771,
      "scale language models": 29136,
      "present indepth analysis": 25536,
      "indepth analysis impact": 14675,
      "neural language model": 22725,
      "work propose use": 35760,
      "build machine learning": 3916,
      "machine learning models": 19349,
      "experiments publicly available": 10473,
      "target domain available": 32049,
      "t5 language model": 31952,
      "tasks text classification": 32527,
      "outperforms strong baselines": 23858,
      "fluent natural language": 11636,
      "achieve good performance": 965,
      "second main contribution": 29324,
      "create synthetic data": 6355,
      "synthetic data improve": 31852,
      "nlp machine learning": 22937,
      "training common practice": 33451,
      "data boost performance": 6628,
      "using synthetic data": 34924,
      "generate synthetic data": 12326,
      "convolutional neural networks": 6125,
      "data improve performance": 6732,
      "performance natural language": 24688,
      "finetune pretrained gpt2": 11299,
      "pretrained gpt2 transformer": 25651,
      "gpt2 transformer model": 12961,
      "model generate synthetic": 20544,
      "natural language explanations": 22516,
      "language model perform": 16186,
      "demonstrated outstanding performance": 7537,
      "performance nlp tasks": 24693,
      "pretrained roberta gpt2": 25744,
      "improve performance particular": 14286,
      "finetuning pretrained language": 11487,
      "pretrained language gpt2": 25656,
      "achieve new stateoftheart": 974,
      "commonsense question answering": 5038,
      "use pretrained language": 34561,
      "paper present novel": 24091,
      "generative models gpt2": 12678,
      "despite recent progress": 7807,
      "models lms t5": 21681,
      "remains largely underexplored": 28002,
      "largely underexplored paper": 17311,
      "paper present study": 24094,
      "present study investigate": 25556,
      "empirical results demonstrate": 9234,
      "best performing models": 3571,
      "analysis reveals models": 1961,
      "dataset publicly available": 7028,
      "autoregressive decoding process": 2938,
      "models t5 gpt2": 22043,
      "source code available": 30549,
      "language models recent": 16683,
      "models recent years": 21909,
      "size pretrained language": 30275,
      "training models scratch": 33569,
      "number taskspecific parameters": 23163,
      "limited computational resources": 17945,
      "downstream tasks experimental": 8690,
      "tasks experimental results": 32322,
      "tens billions parameters": 32717,
      "source code model": 30551,
      "pretraining language understanding": 25808,
      "language understanding generation": 16842,
      "pretrained models achieved": 25717,
      "stateoftheart results various": 30987,
      "results various natural": 28706,
      "processing nlp tasks": 26121,
      "t5 gpt3 shown": 31950,
      "language models improve": 16382,
      "175 billion parameters": 173,
      "propose unified framework": 26581,
      "understanding generation tasks": 34229,
      "generation tasks zeroshot": 12617,
      "tasks zeroshot learning": 32556,
      "learning fewshot learning": 17567,
      "fewshot learning finetuning": 11111,
      "10 billion parameters": 27,
      "results model outperforms": 28646,
      "model outperforms stateoftheart": 20675,
      "outperforms stateoftheart models": 23855,
      "nlp tasks english": 22951,
      "spanish language models": 30595,
      "models pretrained using": 21848,
      "extractive question answering": 10780,
      "question answering dataset": 27042,
      "based large language": 3186,
      "large language model": 16960,
      "language model t5": 16205,
      "tasks conduct extensive": 32273,
      "conduct extensive experiments": 5607,
      "gpt2 based model": 12874,
      "text generation ability": 32866,
      "autoregressive language model": 2943,
      "experimental results performance": 10404,
      "language models complex": 16282,
      "models complex tasks": 21076,
      "relatively small number": 27893,
      "small number examples": 30362,
      "language models small": 16707,
      "training machine learning": 33559,
      "transformerbased pretrained language": 33766,
      "conventional nlp tasks": 6077,
      "language models performance": 16650,
      "models lms exhibit": 21673,
      "human sentence processing": 13863,
      "model best model": 20397,
      "named entity recognition": 22486,
      "entity recognition ner": 9649,
      "significant progress recent": 30014,
      "progress recent years": 26227,
      "stateoftheart sota models": 30993,
      "task aims generate": 32077,
      "facilitate research task": 10845,
      "publicly traded companies": 26867,
      "language model achieving": 16115,
      "achieve sota results": 988,
      "language models financial": 16343,
      "widelyused pretrained language": 35581,
      "holtzman et al": 13748,
      "et al 2020": 9795,
      "hidden states gpt2": 13540,
      "finegrained human annotations": 11274,
      "datasets demonstrate superior": 7092,
      "demonstrate superior performance": 7504,
      "superior performance sota": 31653,
      "processing nlp domain": 26117,
      "performance downstream tasks": 24575,
      "large number parameters": 17253,
      "despite superior performance": 7820,
      "general language understanding": 12174,
      "language understanding evaluation": 16841,
      "evaluation benchmark tasks": 9925,
      "models pretrained language": 21841,
      "wide range natural": 35557,
      "nlp community existing": 22927,
      "existing works focus": 10328,
      "achieve better performance": 948,
      "recently emerged effective": 27594,
      "emerged effective method": 9158,
      "adapting pretrained language": 1216,
      "tasks paper investigate": 32443,
      "natural language utterances": 22576,
      "conduct ablation studies": 5583,
      "different model scales": 8107,
      "increasing model scale": 14623,
      "models like gpt3": 21454,
      "like gpt3 t5": 17870,
      "gpt3 t5 research": 13008,
      "new model architectures": 22824,
      "generalization language models": 12219,
      "language models computational": 16283,
      "recent years pretrained": 27579,
      "years pretrained language": 35896,
      "models bert gpt2": 21018,
      "shown promising results": 29909,
      "neural network architectures": 22740,
      "test set compared": 32787,
      "range downstream tasks": 27193,
      "models deployed resourceconstrained": 21141,
      "proposed framework dubbed": 26600,
      "parameter efficient finetuning": 24180,
      "approach extensive experiments": 2280,
      "backbones bert roberta": 3059,
      "bert roberta gpt2": 3529,
      "roberta gpt2 dozens": 28918,
      "gpt2 dozens datasets": 12885,
      "achieving comparable performance": 1089,
      "language model finetuning": 16145,
      "finetuning language models": 11426,
      "respect input length": 28452,
      "context paper propose": 5906,
      "current pretrained language": 6525,
      "fraction computational cost": 11817,
      "approach using gpt2": 2357,
      "proposed model achieves": 26613,
      "language models data": 16296,
      "data augmentation techniques": 6612,
      "language models shown": 16702,
      "data augmentation text": 6613,
      "improve classification performance": 14258,
      "outperform competitive baselines": 23769,
      "text generation using": 32885,
      "current language models": 6501,
      "language models generate": 16358,
      "models generate highquality": 21308,
      "generate highquality text": 12287,
      "models lstm transformer": 21693,
      "training neural network": 33572,
      "neural network models": 22748,
      "neural networks generalize": 22753,
      "reduce computational cost": 27703,
      "existing methods struggle": 10295,
      "language models catastrophic": 16274,
      "models catastrophic forgetting": 21049,
      "machine learning shifting": 19352,
      "models trained selfsupervised": 22074,
      "large amounts data": 16928,
      "large number tasks": 17254,
      "raises important question": 27168,
      "pretrained models t5": 25728,
      "systems use large": 31922,
      "large neural networks": 17251,
      "neural networks require": 22754,
      "modern language models": 22160,
      "models bert t5": 21020,
      "extensive experiments different": 10697,
      "models trained english": 22067,
      "introduce novel method": 15530,
      "novel method called": 23096,
      "static word embeddings": 31014,
      "outperforms models comparable": 23835,
      "models comparable size": 21070,
      "training large language": 33544,
      "language models new": 16631,
      "make code models": 19455,
      "code models publicly": 4790,
      "models publicly available": 21879,
      "publicly available efficient": 26855,
      "adaptation pretrained language": 1188,
      "language models remarkable": 16693,
      "remarkable success large": 28057,
      "success large language": 31515,
      "language models driven": 16314,
      "models trained massive": 22070,
      "adaptation diverse domains": 1178,
      "method based observation": 19885,
      "frozen pretrained language": 11940,
      "paper proposes efficient": 24122,
      "inference computational cost": 14767,
      "wide range inference": 35554,
      "higher transformer layers": 13612,
      "inference latency experimental": 14787,
      "latency experimental results": 17404,
      "text generation tasks": 32884,
      "benchmarks like glue": 3455,
      "language models llms": 16425,
      "generation recent years": 12591,
      "seq2seq language model": 29590,
      "texttotext language models": 33014,
      "language models structured": 16714,
      "question answering knowledge": 27046,
      "paper overcome limitation": 24084,
      "achieves stateoftheart performance": 1072,
      "zeroshot fewshot learning": 35973,
      "pruning toxicity bias": 26817,
      "language models test": 16729,
      "knowledge distillation pruning": 15836,
      "generate humanlike text": 12289,
      "language models end": 16321,
      "language processing models": 16788,
      "machine learning ml": 19346,
      "analysis neural networks": 1945,
      "tasks prior work": 32457,
      "computer vision cv": 5506,
      "large pretrained transformers": 17270,
      "data model size": 6775,
      "nlp models including": 22939,
      "models including gpt2": 21380,
      "large model pretraining": 17232,
      "higher training throughput": 13610,
      "neural network training": 22750,
      "layer pretrained model": 17430,
      "pretrained model approach": 25712,
      "finetuning pretrained model": 11491,
      "learning models large": 17609,
      "language models llm": 16422,
      "graph neural networks": 13228,
      "language models building": 16268,
      "capable language models": 4110,
      "experiments t5 bert": 10489,
      "code demo available": 4747,
      "language models parameterefficient": 16646,
      "achieve superior performances": 997,
      "language understanding benchmarks": 16839,
      "achieved remarkable success": 1015,
      "model performance compared": 20697,
      "code publicly available": 4803,
      "language generation nlg": 16089,
      "gpt2 generated texts": 12895,
      "training data increase": 33479,
      "data source code": 6871,
      "language models various": 16746,
      "structures neural language": 31233,
      "specific language model": 30701,
      "recurrent neural network": 27682,
      "neural network rnn": 22749,
      "language models transformer": 16738,
      "language models novel": 16636,
      "language models improving": 16385,
      "text generation various": 32886,
      "tasks language models": 32388,
      "completion language models": 5260,
      "zhou et al": 36004,
      "et al 2021": 9796,
      "chen et al": 4530,
      "language model outperforms": 16181,
      "model outperforms gpt2": 20673,
      "al 2019 gpt3": 1683,
      "2019 gpt3 brown": 224,
      "gpt3 brown et": 12984,
      "brown et al": 3898,
      "model code models": 20425,
      "language models deep": 16297,
      "models deep learning": 21128,
      "deep learning dl": 7329,
      "alzheimers disease ad": 1864,
      "ability generalize small": 672,
      "publicly available research": 26861,
      "model parameters directly": 20690,
      "propose novel method": 26551,
      "data widely used": 6914,
      "language models language": 16399,
      "generalization natural language": 12221,
      "processing nlp algorithms": 26115,
      "remains significant challenge": 28013,
      "significant challenge paper": 29965,
      "paper addresses issue": 24006,
      "transformerbased language models": 33749,
      "language models scale": 16698,
      "hundreds billions parameters": 13944,
      "open source available": 23422,
      "training large neural": 33548,
      "address issues propose": 1339,
      "models lms shown": 21679,
      "knowledge pretraining corpora": 15890,
      "performance fewshot scenarios": 24598,
      "alleviates exposure bias": 1794,
      "text generation paper": 32879,
      "language model introduce": 16158,
      "language model trained": 16207,
      "model publicly available": 20739,
      "training evaluation code": 33514,
      "code model weights": 4783,
      "language models successfully": 16720,
      "zero fewshot learning": 35938,
      "models paper introduces": 21794,
      "models 13 billion": 20921,
      "billion 13 billion": 3714,
      "13 billion parameters": 108,
      "billion parameters trained": 3719,
      "low resource languages": 19275,
      "multilingual tasks including": 22333,
      "tasks including classification": 32365,
      "models follow instructions": 21291,
      "despite order magnitude": 7799,
      "order magnitude smaller": 23677,
      "requires significant human": 28262,
      "significant human effort": 29985,
      "paper propose conversational": 24108,
      "user simulator called": 34673,
      "automated natural language": 2871,
      "language generation metrics": 16085,
      "capable providing accurate": 4119,
      "bert language models": 3515,
      "social media platforms": 30425,
      "language models present": 16660,
      "using masked language": 34837,
      "masked language modelling": 19615,
      "generative transformer model": 12711,
      "model capable generating": 20411,
      "deep learning approach": 7325,
      "berts masked language": 3547,
      "masked language modeling": 19612,
      "language modeling mlm": 16221,
      "controlled text generation": 6066,
      "generation tasks demonstrate": 12614,
      "prompting large language": 26382,
      "language model llm": 16165,
      "question answering natural": 27048,
      "answering natural language": 2066,
      "challenge natural language": 4320,
      "processing nlp systems": 26120,
      "machine translation mt": 19360,
      "human evaluation results": 13807,
      "similar model trained": 30110,
      "prompt engineering paper": 26323,
      "introduce new benchmark": 15520,
      "diverse tasks datasets": 8467,
      "translation summarization question": 33851,
      "summarization question answering": 31623,
      "model better results": 20399,
      "language models evaluation": 16327,
      "conduct comprehensive empirical": 5591,
      "comprehensive empirical study": 5366,
      "performance existing stateoftheart": 24590,
      "models significantly outperform": 21976,
      "new stateoftheart performance": 22849,
      "models source code": 21993,
      "models including t5": 21384,
      "language inference nli": 16095,
      "language models right": 16697,
      "set nlp tasks": 29700,
      "propose novel algorithm": 26544,
      "benchmark datasets various": 3375,
      "models bart t5": 21006,
      "bart t5 gpt3": 3110,
      "demonstrate large language": 7467,
      "perform ablation studies": 24470,
      "highlight transformative potential": 13638,
      "language models chatgpt": 16276,
      "field natural language": 11144,
      "endtoend speech recognition": 9441,
      "yield good performance": 35912,
      "tackle challenges propose": 31996,
      "glancing language model": 12788,
      "language model glm": 16148,
      "word error rate": 35639,
      "comparable performance stateoftheart": 5087,
      "generalpurpose pretrained language": 12258,
      "new synthetic data": 22854,
      "issue propose knowledge": 15662,
      "data augmentation model": 6611,
      "language model pretrained": 16191,
      "unified texttotext format": 34340,
      "training objectives different": 33579,
      "best knowledge attempt": 3560,
      "training data augmentation": 33468,
      "extensive experiments synthetic": 10702,
      "models bert albert": 21016,
      "model pretraining finetuning": 20724,
      "finetuning downstream tasks": 11392,
      "variety nlp tasks": 35069,
      "achieve superior performance": 996,
      "college entrance examination": 4941,
      "pretraining natural language": 25825,
      "remarkable success natural": 28060,
      "success natural language": 31520,
      "showcase superior performance": 29842,
      "superior performance compared": 31650,
      "largescale natural language": 17368,
      "text generation model": 32876,
      "challenging task demands": 4399,
      "language model generation": 16147,
      "language models task": 16727,
      "results reveal current": 28673,
      "language models struggle": 16715,
      "representation linguistic phenomena": 28143,
      "pretrained transformerbased language": 25768,
      "language models widely": 16756,
      "models widely used": 22130,
      "understanding nlu natural": 34255,
      "nlu natural language": 22972,
      "training language models": 33540,
      "language models demonstrate": 16298,
      "financial sentiment analysis": 11221,
      "deep learning techniques": 7333,
      "stateoftheart models like": 30958,
      "batch size learning": 3288,
      "size learning rate": 30260,
      "language models infer": 16391,
      "work introduce novel": 35725,
      "representations pretrained language": 28171,
      "language models specifically": 16711,
      "pretrained bert gpt2": 25631,
      "bert gpt2 language": 3510,
      "gpt2 language models": 12911,
      "language models encoder": 16320,
      "natural language datasets": 22512,
      "automatically generate highquality": 2913,
      "widely used datasets": 35577,
      "demonstrate effectiveness method": 7447,
      "different context lengths": 8059,
      "model achieves best": 20347,
      "general language modeling": 12172,
      "language modeling ability": 16215,
      "closedbook question answering": 4684,
      "question answering datasets": 27043,
      "tasks summarization machine": 32520,
      "summarization machine translation": 31619,
      "powered large language": 25333,
      "study shed light": 31398,
      "causal language models": 4243,
      "language models general": 16356,
      "directions future research": 8229,
      "train evaluate models": 33362,
      "problems deep learning": 26024,
      "task use pretrained": 32207,
      "neural network model": 22747,
      "improves model performance": 14385,
      "model performance significantly": 20706,
      "best model outperforms": 3564,
      "outperforms current stateoftheart": 23817,
      "current stateoftheart sota": 6537,
      "trained natural language": 33415,
      "model achieves stateoftheart": 20350,
      "large models nlp": 17238,
      "models nlp tasks": 21755,
      "benefit using large": 3485,
      "using large language": 34808,
      "llms 100 billion": 18399,
      "100 billion parameters": 35,
      "pretrained models scale": 25727,
      "efficient finetuning methods": 9037,
      "language model paper": 16184,
      "recently large language": 27607,
      "deep learning based": 7327,
      "2022 shared task": 231,
      "language models substantially": 16718,
      "prohibitively expensive motivating": 26241,
      "translation natural language": 33838,
      "understanding nlu tasks": 34258,
      "improve performance downstream": 14283,
      "transformers shown remarkable": 33796,
      "shown remarkable success": 29915,
      "performance various tasks": 24808,
      "transformerbased text generation": 33769,
      "transformer models generative": 33733,
      "models generative pretrained": 21316,
      "pretrained transformer gpt": 25757,
      "achieved remarkable performance": 1014,
      "performance text generation": 24784,
      "generation natural language": 12560,
      "performance significantly degrades": 24753,
      "significantly degrades generation": 30045,
      "generation paper present": 12567,
      "xilinx alveo u280": 35872,
      "high bandwidth memory": 13553,
      "bandwidth memory hbm": 3091,
      "models llms training": 21652,
      "recent large language": 27524,
      "models llms demonstrated": 21497,
      "llms demonstrated remarkable": 18520,
      "outperform larger models": 23779,
      "larger language models": 17322,
      "llms significantly outperform": 18951,
      "transformer language model": 33724,
      "given natural language": 12757,
      "natural language prompt": 22560,
      "match exceed performance": 19642,
      "models llms shown": 21629,
      "shown exceptional performance": 29876,
      "exceptional performance variety": 10171,
      "variety natural language": 35066,
      "natural language tasks": 22568,
      "llms indepth analysis": 18716,
      "understanding llms pretrained": 34248,
      "natural language corpora": 22510,
      "compared models trained": 5154,
      "compared previous best": 5163,
      "time memory complexity": 33135,
      "outperforms prior methods": 23848,
      "offtheshelf pretrained language": 23332,
      "data experimental results": 6703,
      "achieves significant improvement": 1063,
      "work focuses simple": 35714,
      "zeroshot fewshot settings": 35976,
      "language models finetuning": 16345,
      "language models collection": 16280,
      "improve model performance": 14276,
      "model performance generalization": 20700,
      "performance generalization unseen": 24612,
      "generalization unseen tasks": 12229,
      "tasks paper explore": 32442,
      "data instruction finetuning": 6740,
      "stateoftheart performance benchmarks": 30970,
      "method improving performance": 19933,
      "evaluation large language": 9966,
      "language models understand": 16740,
      "publicly available pretrained": 26860,
      "language models 13b": 16232,
      "models 13b parameters": 20923,
      "language model train": 16206,
      "language models increasingly": 16388,
      "openaccess multilingual language": 23430,
      "semiparametric language models": 29501,
      "number model parameters": 23152,
      "multiple natural language": 22404,
      "paper develop novel": 24034,
      "semiparametric language model": 29500,
      "language model architecture": 16119,
      "texttotext language model": 33013,
      "superior zeroshot performance": 31659,
      "performance unseen tasks": 24791,
      "smaller model scale": 30382,
      "model scale compared": 20765,
      "leveraging pretrained models": 17793,
      "models recently gained": 21911,
      "models long short": 21688,
      "long short term": 19180,
      "short term memory": 29817,
      "leverage attention mechanism": 17745,
      "pretrained causal language": 25634,
      "model downstream task": 20474,
      "parameter language model": 24185,
      "significant computational resources": 29970,
      "carbon footprint ml": 4161,
      "future research directions": 12045,
      "propose novel learning": 26550,
      "helps language models": 13525,
      "language models better": 16262,
      "models better understand": 21025,
      "using language model": 34805,
      "language model components": 16127,
      "absolute f1 points": 763,
      "large neural language": 17247,
      "synthetic data generation": 31851,
      "generation method based": 12546,
      "multilingual language model": 22311,
      "language model large": 16159,
      "model large language": 20603,
      "shown able perform": 29868,
      "natural language instructions": 22529,
      "led widespread adoption": 17689,
      "language model designed": 16132,
      "achieves competitive performance": 1043,
      "competitive performance wide": 5227,
      "performance wide variety": 24815,
      "multitask prompted finetuning": 22454,
      "facilitate future research": 10841,
      "release models code": 27913,
      "models llms led": 21572,
      "understanding generation abilities": 34228,
      "increasing model size": 14624,
      "size solution propose": 30286,
      "models shown great": 21970,
      "shown great performance": 29880,
      "great performance tasks": 13253,
      "shown improve performance": 29890,
      "performance various nlp": 24806,
      "various nlp tasks": 35131,
      "nlp tasks just": 22956,
      "known incontext learning": 15936,
      "semantic parsing tasks": 29463,
      "tasks incontext learning": 32370,
      "open source model": 23425,
      "underlying language model": 34156,
      "previous supervised stateoftheart": 25888,
      "adapting large language": 1209,
      "finetuning large language": 11428,
      "language models different": 16310,
      "methods reduce number": 20085,
      "conditional language modeling": 5570,
      "language modeling objective": 16222,
      "leverage large language": 17753,
      "language models diverse": 16311,
      "landscape large language": 16024,
      "performance does scale": 24572,
      "llms like gpt": 18759,
      "gpt2 gpt3 models": 12902,
      "models paper examines": 21791,
      "domains using dataset": 8644,
      "largelanguage models llms": 17305,
      "performance smaller models": 24757,
      "smaller models using": 30391,
      "llama2 mpt falcon": 18190,
      "model achieves competitive": 20349,
      "model performance finally": 20699,
      "knowledge large language": 15872,
      "models llms trained": 21651,
      "impressive performance diverse": 14242,
      "language models particular": 16648,
      "active vs passive": 1149,
      "results important aspects": 28629,
      "language modeling task": 16226,
      "multilingual language models": 22312,
      "mbert xlmr mt5": 19713,
      "enabling natural language": 9327,
      "knowledge generative language": 15855,
      "largescale generative language": 17354,
      "language models glms": 16365,
      "secure multiparty computation": 29347,
      "present novel approach": 25545,
      "novel approach generating": 23060,
      "generation task using": 12612,
      "use training data": 34577,
      "gpt2 model able": 12917,
      "previous research focused": 25875,
      "multilingual large language": 22314,
      "dataset used train": 7052,
      "large models datasets": 17234,
      "wide range research": 35562,
      "share lessons learned": 29782,
      "training large deep": 33543,
      "deep neural networks": 7342,
      "language models vision": 16749,
      "base large models": 3120,
      "models trained scratch": 22073,
      "language models training": 16734,
      "incontext learning abilities": 14549,
      "models shown perform": 21971,
      "shown perform better": 29903,
      "wide variety tasks": 35567,
      "incontext learning paradigm": 14564,
      "paper investigate hypothesis": 24070,
      "ability large language": 690,
      "language model incontext": 16155,
      "billion parameter language": 3716,
      "number incontext examples": 23144,
      "overall study provides": 23916,
      "perform incontext learning": 24492,
      "incontext learning multilingual": 14563,
      "nlp large language": 22935,
      "tasks named entity": 32424,
      "pretrained sequencetosequence models": 25749,
      "tuning language models": 33986,
      "language models human": 16379,
      "enables pretrained language": 9307,
      "natural language descriptions": 22514,
      "approaches rely vast": 2389,
      "rely vast amounts": 27974,
      "various benchmarks results": 35079,
      "models llms surprisingly": 21646,
      "generating natural language": 12436,
      "natural language reasoning": 22565,
      "question answering qa": 27051,
      "code data prompts": 4735,
      "safety large language": 29049,
      "different llms using": 8099,
      "gpt35 gpt4 showed": 13026,
      "direct preference optimization": 8214,
      "sequence labeling tasks": 29600,
      "target language paper": 32052,
      "language paper present": 16769,
      "leverages large pretrained": 17769,
      "pretrained texttotext language": 25754,
      "language models stateoftheart": 16713,
      "multilingual t5 model": 22331,
      "lack highquality training": 15991,
      "highquality training data": 13704,
      "tasks code data": 32265,
      "code data publicly": 4737,
      "data publicly available": 6821,
      "generative ai models": 12648,
      "large generative models": 16952,
      "models able perform": 20937,
      "pretrained model finetuning": 25715,
      "bert albert roberta": 3496,
      "recent works proposed": 27573,
      "methods solve problem": 20097,
      "work paper propose": 35741,
      "paper propose novel": 24113,
      "datasets experiment results": 7108,
      "experiment results proposed": 10381,
      "textual style transfer": 33040,
      "propose novel task": 26553,
      "pretrained language generation": 25654,
      "language generation models": 16086,
      "effectiveness large language": 8951,
      "evaluation language models": 9964,
      "performance various natural": 24802,
      "tasks question answering": 32469,
      "question answering summarization": 27055,
      "summarization large language": 31615,
      "models llms used": 21657,
      "language understanding capabilities": 16840,
      "task paper explore": 32172,
      "number examples prompt": 23142,
      "llms paper demonstrate": 18838,
      "multiple tasks including": 22423,
      "gpt2 model generates": 12919,
      "using realworld datasets": 34898,
      "better baseline model": 3595,
      "achieves stateoftheart results": 1073,
      "improve zeroshot generalization": 14303,
      "zeroshot generalization ability": 35978,
      "ability language models": 688,
      "increased model parameters": 14610,
      "generation large language": 12534,
      "finetuned gpt2 model": 11318,
      "gpt2 model trained": 12923,
      "architectures like bert": 2467,
      "computer vision models": 5507,
      "empirical evaluation different": 9221,
      "questionanswering qa datasets": 27086,
      "perform extensive evaluation": 24488,
      "language models fewshot": 16342,
      "finetuned t5 models": 11356,
      "state art large": 30904,
      "models like bert": 21448,
      "like bert gpt": 17846,
      "bert gpt t5": 3508,
      "aspectbased sentiment analysis": 2571,
      "positive negative neutral": 25196,
      "significant performance improvements": 30006,
      "outperforms previous stateoftheart": 23844,
      "previous stateoftheart sota": 25884,
      "strong generalization ability": 31174,
      "qa language models": 26912,
      "stateoftheart language models": 30938,
      "including domain adaptation": 14474,
      "guiding large language": 13361,
      "guide llms generating": 13350,
      "data reinforcement learning": 6830,
      "dialogue response generation": 8019,
      "dialogues multiwoz dataset": 8026,
      "size large language": 30256,
      "language models continue": 16290,
      "reduce computational overhead": 27705,
      "language generation paper": 16090,
      "comprehension natural language": 5348,
      "sequence generation models": 29596,
      "foundation language models": 11794,
      "language models introduce": 16394,
      "language models ranging": 16678,
      "train stateoftheart models": 33377,
      "using publicly available": 34891,
      "publicly available datasets": 26854,
      "release models research": 27914,
      "models research community": 21934,
      "largest language model": 17395,
      "language model explicitly": 16139,
      "available hugging face": 2980,
      "language models plm": 16652,
      "plms gpt2 t5": 25048,
      "language models examine": 16328,
      "models trained large": 22069,
      "language model does": 16134,
      "training data finetuning": 33475,
      "classification semantic segmentation": 4610,
      "analysis question answering": 1954,
      "powerful large language": 25343,
      "natural language question": 22564,
      "question answering using": 27060,
      "question answering kbqa": 27045,
      "dataset code available": 6950,
      "use large language": 34543,
      "models llms chatgpt": 21488,
      "human language processing": 13838,
      "relation extraction given": 27865,
      "model based gpt2": 20387,
      "achieves stateoftheart accuracy": 1071,
      "grammatical error detection": 13208,
      "model paper presents": 20684,
      "language model using": 16210,
      "paper presents detailed": 24097,
      "scaling large language": 29168,
      "language models empirical": 16318,
      "models empirical study": 21193,
      "significantly enhances models": 30050,
      "enhances models performance": 9550,
      "amounts instruction data": 1886,
      "data model performance": 6772,
      "use cases paper": 34517,
      "performance large language": 24645,
      "language models based": 16256,
      "instruction tuning different": 15191,
      "instruction data evaluation": 15144,
      "tasks openended generation": 32436,
      "potential future research": 25256,
      "data large language": 6750,
      "models llms downstream": 21515,
      "models llms revolutionized": 21622,
      "llms revolutionized natural": 18924,
      "revolutionized natural language": 28849,
      "demonstrated impressive capabilities": 7528,
      "capabilities various tasks": 4081,
      "evaluate llms performance": 9846,
      "comparison multiple llms": 5198,
      "conduct comparative analysis": 5588,
      "gpt2 gpt3 chatgpt": 12901,
      "source code data": 30550,
      "deep learning algorithms": 7324,
      "automated machine learning": 2866,
      "machine learning automl": 19345,
      "models llms gpt4": 21547,
      "parameterefficient finetuning large": 24207,
      "language models success": 16719,
      "models llms like": 21574,
      "llms like gpt4": 18763,
      "parameterefficient finetuning peft": 24209,
      "comparable better performance": 5075,
      "llms paper presents": 18841,
      "conduct extensive empirical": 5606,
      "extensive empirical studies": 10683,
      "reasoning commonsense reasoning": 27393,
      "results demonstrate using": 28597,
      "tasks like image": 32400,
      "like image captioning": 17877,
      "feedback large language": 11065,
      "llms like chatgpt": 18757,
      "chatgpt exhibited remarkable": 4466,
      "natural language processingnlp": 22559,
      "models accessible restricted": 20941,
      "barriers new research": 3105,
      "opensource llms llama": 23523,
      "improves translation performance": 14398,
      "language models efficient": 16317,
      "stateoftheart training efficiency": 31002,
      "pretrained models code": 25722,
      "potential large language": 25267,
      "information extraction tasks": 14868,
      "work propose novel": 35756,
      "tasks shows significant": 32502,
      "shows significant improvements": 29937,
      "generated large language": 12368,
      "content recent work": 5872,
      "conduct human evaluation": 5611,
      "text simplification ts": 32943,
      "models outperform models": 21783,
      "paper propose new": 24111,
      "make code publicly": 19458,
      "better instruction following": 3609,
      "instruction following language": 15165,
      "language models chinese": 16277,
      "impact training data": 14141,
      "influence training data": 14841,
      "highquality instruction datasets": 13692,
      "offering valuable insights": 23304,
      "opensource chat models": 23487,
      "proprietary language models": 26638,
      "make model data": 19473,
      "model data code": 20449,
      "data code publicly": 6636,
      "instruction data instruction": 15147,
      "instruction following large": 15166,
      "following large language": 11696,
      "language models crucial": 16293,
      "research field natural": 28313,
      "tuning techniques lora": 34018,
      "llama base model": 18079,
      "language models especially": 16325,
      "especially field chinese": 9736,
      "concerns regarding potential": 5547,
      "llms chatgpt gpt4": 18465,
      "transformed natural language": 33700,
      "artificial general intelligence": 2531,
      "general intelligence agi": 12169,
      "research large language": 28331,
      "language models llama": 16421,
      "capabilities understanding generating": 4076,
      "ability follow instructions": 667,
      "secondary pretraining using": 29332,
      "data finetune model": 6712,
      "enhancing models ability": 9571,
      "experimental results indicate": 10399,
      "proficiency understanding generating": 26183,
      "yield competitive performance": 35910,
      "competitive performance models": 5225,
      "size pretrained models": 30278,
      "open research community": 23419,
      "various downstream tasks": 35090,
      "data filtering process": 6710,
      "bert t5 model": 3534,
      "effective instruction tuning": 8877,
      "select diverse set": 29377,
      "long text generation": 19186,
      "generation models outperform": 12556,
      "outperform 10x larger": 23763,
      "language models instruction": 16393,
      "models instruction tuning": 21401,
      "instruction tuning tasks": 15208,
      "longform question answering": 19206,
      "deep learning code": 7328,
      "critical machine learning": 6390,
      "applications large language": 2160,
      "machine learning systems": 19353,
      "llms perform worse": 18849,
      "lowrank adaptation lora": 19301,
      "datasets large language": 7139,
      "smaller models finetuned": 30388,
      "taskoriented dialog systems": 32219,
      "et al 2023": 9798,
      "using wide range": 34941,
      "wide range models": 35556,
      "given recent success": 12767,
      "encoderdecoder model mt0": 9369,
      "et al 2022": 9797,
      "languages intentionally seen": 16880,
      "model outperforms baseline": 20671,
      "baseline large margin": 3251,
      "findings reveal models": 11250,
      "models training data": 22078,
      "deploying large language": 7640,
      "models llms challenging": 21487,
      "amounts training data": 1891,
      "training data achieve": 33467,
      "data achieve comparable": 6589,
      "achieve comparable performance": 953,
      "achieves better performance": 1035,
      "substantially smaller model": 31488,
      "reduce model size": 27720,
      "dataset release code": 7033,
      "models pretrained large": 21843,
      "finetuned model perform": 11343,
      "suggest language models": 31572,
      "language models learn": 16413,
      "models generate text": 21310,
      "artificial intelligence ai": 2534,
      "language models despite": 16304,
      "prompt tuning simple": 26355,
      "simple efficient method": 30148,
      "efficient method significantly": 9049,
      "method significantly improves": 19974,
      "significantly improves performance": 30062,
      "era large language": 9698,
      "language models gpt3": 16370,
      "unlabeled training data": 34383,
      "pretraining language models": 25807,
      "language models ptlms": 16675,
      "building recent progress": 3927,
      "descriptions large language": 7689,
      "yields best performance": 35922,
      "language models temporal": 16728,
      "natural language nl": 22534,
      "different application domains": 8045,
      "domains paper propose": 8633,
      "models llms multiple": 21588,
      "achieves higher accuracy": 1047,
      "training data compared": 33471,
      "data compared baseline": 6647,
      "graphical user interfaces": 13235,
      "user interfaces guis": 34661,
      "natural language interfaces": 22530,
      "generalpurpose language models": 12249,
      "models finetuned specific": 21278,
      "language models follow": 16351,
      "deployed language models": 7634,
      "leveraging large language": 17784,
      "prediction large language": 25427,
      "models llms produce": 21608,
      "enhancing large language": 9563,
      "advancements large language": 1466,
      "artificial intelligence systems": 2538,
      "closedsource models like": 4689,
      "models like chatgpt": 21451,
      "like chatgpt opensource": 17856,
      "opensource models like": 23531,
      "language models given": 16364,
      "llms various sizes": 19035,
      "including llama alpaca": 14499,
      "language model create": 16129,
      "alignment large language": 1766,
      "end tasks user": 9420,
      "tasks user preferences": 32544,
      "llama language model": 18116,
      "language model finetuned": 16144,
      "model finetuned standard": 20533,
      "reinforcement learning human": 27841,
      "training data including": 33478,
      "instruction tuning data": 15188,
      "produce high quality": 26147,
      "harnessing capabilities large": 13460,
      "capabilities large language": 4032,
      "paper explore prompting": 24047,
      "llm families bloom": 18302,
      "detailed ablation studies": 7832,
      "best knowledge work": 3561,
      "language models acquire": 16241,
      "performance variety language": 24794,
      "paper investigate ability": 24067,
      "models including gpt3": 21381,
      "language models alms": 16247,
      "large pretrained models": 17267,
      "using large pretrained": 34814,
      "generative large language": 12667,
      "llms raises question": 18886,
      "llms revolutionized field": 18923,
      "comes significant computational": 4972,
      "significant computational costs": 29968,
      "computational costs paper": 5463,
      "costs paper propose": 6273,
      "efficient llm inference": 9045,
      "power llms approach": 25325,
      "proprietary large language": 26640,
      "garnered significant attention": 12123,
      "significant attention exceptional": 29961,
      "diverse range tasks": 8455,
      "llama display remarkable": 18093,
      "tuning experimental results": 33978,
      "evaluating large language": 9900,
      "language models spoken": 16712,
      "models spoken language": 22005,
      "spoken language understanding": 30816,
      "models demonstrated strong": 21137,
      "strong language understanding": 31179,
      "language understanding slu": 16849,
      "results smaller models": 28687,
      "ability foundation models": 669,
      "wide range linguistic": 35555,
      "models commonsense knowledge": 21068,
      "models llms directly": 21514,
      "paper shows llms": 24134,
      "carlo tree search": 4185,
      "llm world model": 18381,
      "text generated large": 32862,
      "extensive human evaluation": 10708,
      "strong language model": 31178,
      "paper explores potential": 24050,
      "models llms data": 21493,
      "evaluate effectiveness finetuning": 9829,
      "multilingual models mbert": 22321,
      "models mbert xlmr": 21709,
      "training data generated": 33477,
      "furthermore conduct human": 11990,
      "languages like tamil": 16890,
      "conversational ai systems": 6095,
      "automatic speech recognition": 2895,
      "speech recognition asr": 30788,
      "utilization large language": 34956,
      "bias large language": 3651,
      "ner sentiment analysis": 22679,
      "present systematic study": 25559,
      "comprehensive evaluation large": 5370,
      "language models automatic": 16254,
      "different data sources": 8066,
      "make data code": 19462,
      "hallucination large language": 13379,
      "language models inference": 16392,
      "tasks large language": 32391,
      "capable natural language": 4115,
      "tasks like question": 32402,
      "like question answering": 17893,
      "decoding language models": 7275,
      "families including opt": 10968,
      "models lms powerful": 21676,
      "propose adapt pretrained": 26491,
      "language models capable": 16270,
      "model soft prompts": 20803,
      "opt llama2 models": 23598,
      "reducing inference costs": 27752,
      "extend context window": 10650,
      "llm able perform": 18261,
      "finetuning llms using": 11450,
      "using instruction tuning": 34801,
      "instruction tuning particular": 15205,
      "instruction tuning dataset": 15189,
      "emerges promising solution": 9189,
      "models recent advancements": 21902,
      "recent advancements large": 27492,
      "models llms significantly": 21639,
      "speech processing tasks": 30786,
      "processing tasks including": 26129,
      "fewshot learning techniques": 11114,
      "sota models llms": 30536,
      "provide valuable insights": 26735,
      "semantic textual similarity": 29480,
      "described natural language": 7679,
      "language model evaluation": 16138,
      "diverse natural language": 8442,
      "learning large language": 17593,
      "generation instruction following": 12525,
      "instruction following abilities": 15162,
      "training set containing": 33611,
      "finetune llama7b model": 11292,
      "model needs learn": 20656,
      "performs significantly better": 24855,
      "openended question answering": 23460,
      "question answering fact": 27044,
      "text simplification models": 32942,
      "pretrained multilingual language": 25733,
      "language models reveal": 16696,
      "zeroshot crosslingual transfer": 35965,
      "crosslingual transfer lowresource": 6421,
      "language model finetune": 16143,
      "gap open closed": 12099,
      "lms current methods": 19078,
      "tasks unlike prior": 32540,
      "unlike prior works": 34402,
      "extremescale teacher model": 10808,
      "pretrained lms gpt2": 25709,
      "strong baselines including": 31164,
      "small finetuned models": 30342,
      "new tasks domains": 22857,
      "model weights available": 20866,
      "task machine translation": 32155,
      "language models partially": 16647,
      "models llms acquire": 21483,
      "incontext learning icl": 14555,
      "results provide evidence": 28664,
      "capabilities pretrained language": 4062,
      "language models wide": 16755,
      "pretrained large language": 25698,
      "recent studies ability": 27553,
      "gpt2 empirically demonstrate": 12887,
      "variety language tasks": 35063,
      "excel various natural": 10156,
      "tasks current research": 32282,
      "current research focuses": 6528,
      "llms including gpt3": 18693,
      "demonstrate incontext learning": 7465,
      "incontext learning instruction": 14560,
      "learning instruction tuning": 17589,
      "recent years significant": 27582,
      "language models nlp": 16634,
      "datasets address issue": 7062,
      "positive negative examples": 25195,
      "challenging paper proposes": 4391,
      "especially large language": 9740,
      "transformer language models": 33725,
      "autoregressive text generation": 2955,
      "performance gap small": 24608,
      "training language modeling": 33539,
      "methods fall short": 20036,
      "fall short addressing": 10951,
      "wide range tasks": 35563,
      "models different scales": 21155,
      "models llms study": 21645,
      "language models help": 16377,
      "comparing language models": 5191,
      "parameters language models": 24260,
      "largescale transformer models": 17384,
      "models shown remarkable": 21972,
      "shown remarkable performance": 29914,
      "performance language modelling": 24642,
      "language modelling tasks": 16230,
      "prohibitive training costs": 26238,
      "reduce number parameters": 27722,
      "model downstream tasks": 20475,
      "downstream tasks including": 8696,
      "tasks including language": 32366,
      "model performs similarly": 20710,
      "large foundation models": 16946,
      "surpasses conventional stateoftheart": 31740,
      "zeroshot reasoning benchmarks": 35992,
      "bigbench hard bbh": 3702,
      "shows competitive performance": 29925,
      "large language modelsllms": 17222,
      "using opensource llm": 34867,
      "improving zeroshot performance": 14427,
      "variety downstream tasks": 35061,
      "downstream tasks code": 8688,
      "children language models": 4535,
      "deep language models": 7321,
      "gpt2 models scratch": 12928,
      "models tend learn": 22050,
      "shed new light": 29796,
      "process natural language": 26076,
      "reasoning question answering": 27442,
      "language models using": 16745,
      "entities pretrained language": 9636,
      "questionanswering tasks work": 27091,
      "propose techniques improve": 26573,
      "structured knowledge graphs": 31223,
      "lossless text compression": 19257,
      "language models provide": 16673,
      "comprehensive benchmark study": 5357,
      "encoderdecoder models mt5": 9371,
      "achieve highest performance": 969,
      "language models bloom": 16267,
      "foundation language model": 11793,
      "models llms achieved": 21475,
      "llms achieved great": 18413,
      "achieved great success": 1007,
      "research applications field": 28292,
      "demonstrate effectiveness approach": 7446,
      "model llm based": 20621,
      "based chat assistants": 3143,
      "multilingual pretrained models": 22326,
      "reasoning tasks multilingual": 27458,
      "pretrained model does": 25714,
      "different types tasks": 8155,
      "2023 shared task": 235,
      "ability generative language": 679,
      "stateoftheart models including": 30957,
      "models including alpaca": 21378,
      "automated human evaluation": 2864,
      "translation large language": 33831,
      "recent years large": 27576,
      "years large language": 35892,
      "language models open": 16639,
      "gpt4 metas llama": 13096,
      "metas llama googles": 19864,
      "language models languages": 16403,
      "models work explore": 22134,
      "work explore capabilities": 35705,
      "analysis large language": 1935,
      "large multilingual language": 17242,
      "language models understanding": 16741,
      "language large language": 16106,
      "generative artificial intelligence": 12654,
      "emergence large language": 9170,
      "llms telecom domain": 18999,
      "demonstrate use case": 7509,
      "accuracy gpt2 model": 887,
      "achieves similar performance": 1066,
      "capabilities natural language": 4048,
      "pose significant risks": 25162,
      "opensource code repositories": 23492,
      "based generative pretrained": 3165,
      "generative pretrained transformers": 12699,
      "pretrained transformers gpts": 25772,
      "opensource language models": 23508,
      "potential artificial general": 25241,
      "model language models": 20601,
      "language models investigate": 16395,
      "received little attention": 27479,
      "alignment instruction following": 1764,
      "llms instruction tuning": 18724,
      "plays vital role": 25033,
      "llms human preferences": 18679,
      "performance nonenglish languages": 24695,
      "gpt4 automatic evaluation": 13056,
      "instruction test set": 15181,
      "language models scientific": 16700,
      "gap theory practice": 12113,
      "model size training": 20795,
      "counterfactual data augmentation": 6296,
      "data augmentation method": 6609,
      "impact models performance": 14134,
      "mitigate problem propose": 20257,
      "problem propose novel": 26005,
      "augmentation method generate": 2803,
      "text pretrained language": 32918,
      "language model plm": 16188,
      "model plm t5": 20713,
      "performs better current": 24843,
      "current augmentation methods": 6484,
      "new large language": 22814,
      "language model code": 16126,
      "wide range domains": 35551,
      "computational resources time": 5480,
      "process reduces computational": 26080,
      "reduces computational requirements": 27734,
      "reduces training time": 27741,
      "empowering large language": 9270,
      "llms including gpt4": 18699,
      "including gpt4 llama": 14491,
      "gpt4 llama chat": 13090,
      "finetuning parameterefficient finetuning": 11471,
      "adapt pretrained language": 1169,
      "tasks paper propose": 32444,
      "additional training enables": 1267,
      "instructiontuned large language": 15287,
      "model based llama": 20388,
      "results demonstrate approach": 28587,
      "pretrained transformer gpt4": 25761,
      "code generation machine": 4759,
      "generation machine translation": 12544,
      "models llms capture": 21486,
      "address issue work": 1335,
      "manner experimental results": 19548,
      "demonstrated remarkable potential": 7546,
      "potential natural language": 25282,
      "llms remains significant": 18906,
      "popular offtheshelf llms": 25132,
      "experimental results demonstrate": 10391,
      "results demonstrate superior": 28594,
      "methods extensive experiments": 20032,
      "extensive experiments demonstrate": 10695,
      "publicly available multilingual": 26859,
      "model outperforms previous": 20674,
      "previous stateoftheart models": 25883,
      "performs competitively compared": 24849,
      "nlp tasks large": 22957,
      "models llms typically": 21654,
      "nlp tasks despite": 22949,
      "scale model parameters": 29142,
      "outperforms taskspecific models": 23861,
      "factual consistency language": 10879,
      "model improves various": 20576,
      "gpt35 chatgpt gpt4": 13019,
      "question answering tasks": 27057,
      "exact match em": 10089,
      "llms including gpt2": 18692,
      "research highlights need": 28323,
      "language models generating": 16361,
      "shared task study": 29785,
      "abilities large language": 629,
      "language models providing": 16674,
      "present extensive evaluation": 25532,
      "models including gpt4": 21383,
      "fewshot incontext learning": 11108,
      "using reinforcement learning": 34901,
      "skills large language": 30311,
      "models gpt4 gpt35": 21333,
      "project website available": 26248,
      "language models current": 16295,
      "developments large language": 7981,
      "models llms enabled": 21519,
      "simple general effective": 30151,
      "improve performance language": 14284,
      "language models math": 16616,
      "solve math problems": 30493,
      "problems language models": 26029,
      "long context understanding": 19168,
      "models llms recently": 21611,
      "llms recently achieved": 18895,
      "following natural language": 11699,
      "higher success rate": 13607,
      "models graphtotext generation": 21338,
      "models llms widely": 21664,
      "finetuning llms requires": 11449,
      "llms requires significant": 18913,
      "performance finetuned llm": 24602,
      "models capable generating": 21041,
      "generating fluent coherent": 12424,
      "fluent coherent text": 11634,
      "error analysis reveals": 9709,
      "models struggle understanding": 22017,
      "detect machinegenerated text": 7851,
      "llms shown remarkable": 18944,
      "effectiveness various generaldomain": 8972,
      "various generaldomain natural": 35097,
      "generaldomain natural language": 12192,
      "address challenge introduce": 1309,
      "language models multiple": 16630,
      "learning human feedback": 17576,
      "development large language": 7958,
      "supervised finetuning sft": 31679,
      "finetuning sft reinforcement": 11521,
      "sft reinforcement learning": 29765,
      "human feedback rlhf": 13821,
      "commercial llms chatgpt": 4989,
      "existing opensource llms": 10305,
      "instruction tuning llms": 15201,
      "impressive results various": 14248,
      "performance approach involves": 24525,
      "different model sizes": 8108,
      "llms generate highquality": 18635,
      "achieve performance comparable": 978,
      "achieves new stateoftheart": 1053,
      "new stateoftheart result": 22850,
      "new stateoftheart results": 22851,
      "tasks code available": 32264,
      "reasoning large language": 27416,
      "models llms exhibit": 21525,
      "tasks wide range": 32550,
      "current stateoftheart llms": 6533,
      "highlights need research": 13654,
      "opensource large language": 23510,
      "great success large": 13263,
      "language models computer": 16284,
      "led paradigm shift": 17687,
      "current large language": 6503,
      "recent advances large": 27500,
      "advances large language": 1479,
      "accelerate llm inference": 792,
      "smaller language models": 30377,
      "exhibit strong reasoning": 10230,
      "dense retrieval method": 7611,
      "existing large language": 10284,
      "imbalance training data": 14102,
      "instructionfollowing large language": 15231,
      "source code summarization": 30554,
      "code summarization task": 4820,
      "writing natural language": 35854,
      "semantic similarity metric": 29475,
      "model gpt2 sequence": 20556,
      "average bleu score": 3012,
      "language models optimization": 16641,
      "behavior large language": 3316,
      "supervised finetuning reinforcement": 31676,
      "finetuning reinforcement learning": 11504,
      "natural language specification": 22567,
      "learning ml models": 17606,
      "like bert roberta": 17849,
      "llms focusing llama": 18617,
      "pretraining objective llms": 25828,
      "enhances understanding llms": 9554,
      "gpt35 palm2 llama2": 13034,
      "provide indepth analysis": 26708,
      "model size number": 20793,
      "size number parameters": 30268,
      "pretrained models despite": 25723,
      "llama llama2 models": 18121,
      "number tokens required": 23165,
      "resulting model named": 28558,
      "fewshot learning tasks": 11113,
      "foundational language models": 11804,
      "advanced natural language": 1434,
      "processing nlp research": 26119,
      "reinforcement learning approach": 27840,
      "instructions large language": 15257,
      "models llms present": 21604,
      "instruction tuning standard": 15207,
      "results demonstrate significant": 28593,
      "evaluate ability llms": 9818,
      "based blooms taxonomy": 3141,
      "potential data leakage": 25252,
      "llms comprehensive experiments": 18481,
      "experiments advanced llms": 10418,
      "achieves sota performance": 1068,
      "substantial room improvement": 31477,
      "room improvement especially": 28968,
      "data codes publicly": 6639,
      "codes publicly available": 4855,
      "language models advent": 16245,
      "models advent large": 20970,
      "advent large language": 1510,
      "revolutionized field natural": 28845,
      "language processing enabling": 16780,
      "significant progress various": 30016,
      "powerful models knowledge": 25348,
      "tasks paper proposes": 32445,
      "language models focus": 16349,
      "low rank adaptation": 19273,
      "dense passage retrieval": 7609,
      "average f1 score": 3016,
      "user study comparing": 34676,
      "gpt2 model model": 12920,
      "federated learning fl": 11051,
      "parameterefficient training methods": 24215,
      "orders magnitude faster": 23682,
      "use large transformerbased": 34546,
      "large transformerbased models": 17282,
      "transformerbased models bert": 33761,
      "models bert gpt": 21017,
      "significant advancements natural": 29955,
      "advancements natural language": 1471,
      "models computationally expensive": 21084,
      "models range natural": 21886,
      "deep reinforcement learning": 7344,
      "reinforcement learning rl": 27845,
      "learning rl based": 17647,
      "vision language models": 35303,
      "paper proposes novel": 24125,
      "t5 model generate": 31955,
      "integrating large language": 15332,
      "efficacy proposed approach": 8992,
      "intelligence large language": 15359,
      "recent progress large": 27542,
      "progress large language": 26213,
      "development artificial intelligence": 7950,
      "intelligence ai based": 15353,
      "dataset evaluate effectiveness": 6982,
      "evaluate effectiveness llms": 9830,
      "chainofthought cot think": 4299,
      "cot think stepbystep": 6285,
      "evaluation popular llms": 9991,
      "models using methods": 22104,
      "performance improvements compared": 24628,
      "models different sizes": 21156,
      "language understanding reasoning": 16848,
      "unified language model": 34331,
      "language model work": 16212,
      "openais gpt series": 23441,
      "marked significant advancement": 19594,
      "vast amounts text": 35182,
      "amounts text data": 1889,
      "understanding generating humanlike": 34225,
      "generating humanlike text": 12429,
      "diverse range topics": 8456,
      "llms exploring potential": 18594,
      "stateoftheart llms gpt35": 30946,
      "llms gpt35 gpt4": 18651,
      "prompt engineering techniques": 26324,
      "llms data preprocessing": 18504,
      "accuracy f1 score": 880,
      "study underscores promise": 31404,
      "address problems propose": 1352,
      "including llama bert": 14500,
      "demonstrating superiority existing": 7593,
      "model size findings": 20790,
      "llms code available": 18469,
      "framework pretraining finetuning": 11889,
      "models like t5": 21460,
      "address challenge present": 1311,
      "efficient pretraining finetuning": 9055,
      "language modelling research": 16229,
      "technical report large": 32608,
      "report large language": 28118,
      "progress opensource llms": 26225,
      "7b parameter models": 541,
      "parameter models 8k": 24192,
      "better results compared": 3626,
      "results compared stateoftheart": 28582,
      "modeling tasks shows": 20909,
      "llms achieved remarkable": 18417,
      "nlp multimodal tasks": 22941,
      "existing evaluations focus": 10273,
      "experimental results model": 10402,
      "models despite impressive": 21146,
      "despite impressive capabilities": 7789,
      "impressive capabilities large": 14232,
      "retrieved external knowledge": 28775,
      "llama family models": 18101,
      "investigate question introduce": 15595,
      "llms gpt4 palm": 18660,
      "gpt4 palm llama": 13109,
      "fake news detection": 10935,
      "news detection using": 22881,
      "finetuned large language": 11325,
      "model paper considers": 20682,
      "paper considers possibility": 24028,
      "finetuning llama large": 11441,
      "llama large language": 18118,
      "finetuning peftlora based": 11479,
      "peftlora based approach": 24441,
      "based approach used": 3134,
      "approach used study": 2352,
      "used study model": 34625,
      "study model finetuned": 31363,
      "model finetuned following": 20527,
      "finetuned following tasks": 11313,
      "following tasks analysing": 11705,
      "tasks analysing text": 32242,
      "extracting named entities": 10760,
      "sentiments obtained results": 29579,
      "obtained results finetuned": 23258,
      "results finetuned llama": 28614,
      "finetuned llama model": 11330,
      "llama model perform": 18129,
      "extracted sentiments named": 10753,
      "sentiments named entities": 29575,
      "named entities considered": 22482,
      "entities considered predictive": 9631,
      "considered predictive features": 5719,
      "predictive features supervised": 25450,
      "features supervised machine": 11040,
      "supervised machine learning": 31685,
      "family large language": 10976,
      "smaller transformerbased language": 30400,
      "model produce coherent": 20729,
      "produce coherent english": 26141,
      "billion parameter model": 3717,
      "models llms generate": 21539,
      "enhance learning process": 9515,
      "common sense reasoning": 5014,
      "reasoning natural language": 27430,
      "complex reasoning tasks": 5290,
      "publicly available llms": 26858,
      "models like llama": 21458,
      "demonstrate significant potential": 7495,
      "prominent llms including": 26269,
      "llms including gpt35": 18694,
      "including gpt35 gpt4": 14486,
      "language models exhibit": 16330,
      "propose novel approach": 26545,
      "gpt2 pretrained language": 12938,
      "language model corpus": 16128,
      "astronomy large language": 2673,
      "language models excel": 16329,
      "specialized domains like": 30672,
      "bridge gap introduce": 3865,
      "model finetuned llama2": 20531,
      "finetuned llama2 using": 11335,
      "causal language modeling": 4242,
      "models despite having": 21145,
      "despite having significantly": 7784,
      "exploring large language": 10619,
      "models llms gpt": 21543,
      "findings suggest llms": 11258,
      "chatgpt llama2 models": 4488,
      "accelerating large language": 800,
      "models llms need": 21591,
      "proposed method requires": 26609,
      "method requires additional": 19969,
      "generation language models": 12532,
      "models continue advance": 21108,
      "evaluate capabilities language": 9821,
      "capabilities language models": 4030,
      "address gap propose": 1326,
      "gap propose novel": 12106,
      "traditional chinese benchmarks": 33344,
      "offer comprehensive evaluation": 23288,
      "comprehensive evaluation framework": 5369,
      "assessment language models": 2617,
      "language models capabilities": 16269,
      "performance comparable gpt35": 24549,
      "models llms excel": 21522,
      "llms excel various": 18576,
      "carefully crafted prompts": 4175,
      "propose novel framework": 26549,
      "natural language expressions": 22518,
      "language processing capabilities": 16777,
      "closed opensource llms": 4680,
      "opensource llms including": 23521,
      "covering language understanding": 6326,
      "prompts existing methods": 26417,
      "automatic prompt generation": 2891,
      "foundational large language": 11806,
      "opendomain question answering": 23451,
      "used tune llms": 34634,
      "power large language": 25321,
      "struggle tasks require": 31245,
      "perform comprehensive evaluation": 24480,
      "language models dynamic": 16315,
      "inference large language": 14782,
      "generative nlp tasks": 12684,
      "making large language": 19508,
      "alpaca dataset instruction": 1827,
      "dataset instruction following": 7005,
      "memory usage inference": 19835,
      "li et al": 17809,
      "method generating text": 19927,
      "text language models": 32901,
      "generative model inference": 12674,
      "large gpu memory": 16954,
      "gpu memory consumption": 13174,
      "reduce gpu memory": 27712,
      "gpu memory footprint": 13175,
      "main bottleneck generative": 19388,
      "memory bandwidth bottleneck": 19803,
      "significantly outperforms stateoftheart": 30079,
      "demonstrated remarkable performance": 7544,
      "examples natural language": 10138,
      "technical report present": 32611,
      "outperforms opensource models": 23841,
      "opensource models similar": 23534,
      "benchmarks like mmlu": 3457,
      "language models process": 16665,
      "evaluation natural language": 9980,
      "high error rates": 13567,
      "chatgpt gpt4 bard": 4479,
      "questions large language": 27118,
      "results demonstrate effectiveness": 28588,
      "enhance capabilities large": 9506,
      "language models educational": 16316,
      "language models research": 16695,
      "best configuration outperforms": 3557,
      "13b model trained": 128,
      "number training tokens": 23172,
      "training tokens significant": 33637,
      "small language models": 30348,
      "text style transfer": 32952,
      "impact large language": 14128,
      "high deployment costs": 13565,
      "evaluation text generation": 10021,
      "text generation quality": 32882,
      "machine translation large": 19356,
      "field machine translation": 11140,
      "machine translation recent": 19364,
      "translation recent work": 33849,
      "translation nmt systems": 33840,
      "cases large language": 4201,
      "models llms emerged": 21517,
      "outputs paper study": 23897,
      "incontext learning finetuning": 14553,
      "provides valuable insights": 26768,
      "paper proposes comprehensive": 24121,
      "various benchmarks including": 35078,
      "codes data models": 4852,
      "models llms represent": 21620,
      "way interact computers": 35437,
      "evaluate performance llms": 9855,
      "llms based 13": 18444,
      "evaluate effectiveness models": 9831,
      "performing models achieved": 24836,
      "models achieved accuracy": 20948,
      "calculations large language": 3969,
      "language models highquality": 16378,
      "model finetuned llama": 20530,
      "code models datasets": 4789,
      "autoregressive large language": 2948,
      "model trained sentence": 20837,
      "reversal curse finetuning": 28821,
      "chatgpt gpt35 gpt4": 4476,
      "paper explore capabilities": 24046,
      "nlp tasks text": 22964,
      "tasks text summarization": 32529,
      "text classification sentiment": 32826,
      "classification sentiment analysis": 4612,
      "performance opensource llms": 24704,
      "opensource llms like": 23522,
      "reasoning ability llms": 27374,
      "ability llms large": 700,
      "llms large language": 18744,
      "remarkable performance wide": 28049,
      "performance wide range": 24814,
      "pose challenges practical": 25159,
      "challenges practical deployment": 4370,
      "smaller models distillation": 30385,
      "models specifically tailored": 22003,
      "scientific tabletotext generation": 29258,
      "smaller models experimental": 30386,
      "models experimental results": 21236,
      "distilled data achieves": 8352,
      "significant improvement compared": 29990,
      "significantly better performance": 30036,
      "speedup modern hardware": 30802,
      "ability use knowledge": 727,
      "language models ability": 16234,
      "methods large language": 20058,
      "models llms gained": 21535,
      "language models slms": 16706,
      "limited labeled data": 17953,
      "remarkable performance gain": 28047,
      "language models solving": 16708,
      "llms shown promise": 18943,
      "diverse question types": 8452,
      "especially smaller models": 9748,
      "smaller models like": 30390,
      "models like llama2": 21459,
      "improvements natural language": 14360,
      "models fall short": 21261,
      "chain thoughts prompting": 4294,
      "language models generative": 16363,
      "llms demonstrated impressive": 18516,
      "impressive performance various": 14243,
      "solving math word": 30511,
      "math word problems": 19676,
      "primary aim research": 25916,
      "training large models": 33547,
      "tasks results suggest": 32491,
      "results suggest models": 28693,
      "metric text generation": 20127,
      "guided natural language": 13354,
      "natural language instruction": 22528,
      "approach large language": 2306,
      "different model families": 8106,
      "context downstream tasks": 5887,
      "significantly boost performance": 30038,
      "sequence sequence models": 29607,
      "language models reasoning": 16682,
      "facilitate comprehensive evaluation": 10836,
      "reasoning capabilities large": 27384,
      "extensive evaluation using": 10686,
      "using popular llms": 34876,
      "popular llms gpt4": 25125,
      "llms gpt4 llama2": 18657,
      "fewshot learning scenarios": 11112,
      "findings indicate models": 11239,
      "reasoning abilities llms": 27372,
      "providing nuanced understanding": 26781,
      "data recent advancements": 6827,
      "recent advancements llms": 27495,
      "llms demonstrated potential": 18519,
      "reasoning paths using": 27432,
      "opensource llm series": 23518,
      "method achieves stateoftheart": 19870,
      "explore large language": 10588,
      "using proposed method": 34887,
      "recent success large": 27560,
      "gained significant attention": 12066,
      "zero fewshot generalization": 35937,
      "capabilities opensource llms": 4058,
      "token classification tasks": 33186,
      "substantially outperforms llms": 31486,
      "llms recently showcased": 18897,
      "recently showcased remarkable": 27623,
      "processing tasks diverse": 26127,
      "tasks diverse domains": 32300,
      "llms paper propose": 18842,
      "paper propose framework": 24109,
      "textdavinci003 gpt35turbo gpt4": 32982,
      "gpt4 llama27b llama213b": 13094,
      "bypass safety alignment": 3949,
      "games large language": 12081,
      "language models systematically": 16724,
      "evaluate various llms": 9870,
      "significant differences performance": 29979,
      "consistency language models": 5733,
      "llms fall short": 18602,
      "based multiagent collaboration": 3199,
      "evaluate capabilities llms": 9823,
      "language models represent": 16694,
      "models llms sparked": 21642,
      "language modeling question": 16224,
      "modeling question answering": 20906,
      "costs large language": 6270,
      "language model inference": 16156,
      "models llms exploded": 21527,
      "llms exploded popularity": 18590,
      "recent stateoftheart llm": 27551,
      "knowledge work study": 15922,
      "remains open problem": 28006,
      "language models contain": 16289,
      "language model demonstrate": 16131,
      "tasks finetuning language": 32337,
      "data training evaluation": 6897,
      "zeroshot chain thought": 35961,
      "chain thought prompting": 4292,
      "language models good": 16367,
      "remarkable success wide": 28063,
      "work propose new": 35754,
      "propose new benchmark": 26537,
      "new benchmark termed": 22782,
      "finetuning experimental results": 11401,
      "llms achieved impressive": 18415,
      "answer complex questions": 2042,
      "advancements artificial intelligence": 1461,
      "advanced large language": 1426,
      "llama shown great": 18144,
      "shown great potential": 29882,
      "generative ai genai": 12647,
      "llms demonstrates significant": 18530,
      "language model capabilities": 16125,
      "single attention head": 30199,
      "models llms pretrained": 21606,
      "datasets work introduce": 7192,
      "smaller models bloomz": 30384,
      "joint entity relation": 15713,
      "using single model": 34913,
      "corresponding entity relation": 6227,
      "gap introduce new": 12092,
      "method outperforms existing": 19952,
      "benchmarking large language": 3424,
      "language models augmented": 16252,
      "task natural language": 32165,
      "learning techniques work": 17663,
      "work paves way": 35745,
      "text data augmentation": 32842,
      "generate high quality": 12283,
      "finetuning language model": 11425,
      "application machine learning": 2136,
      "demonstrated remarkable capabilities": 7543,
      "remarkable capabilities natural": 28033,
      "various domains including": 35088,
      "llms achieve similar": 18411,
      "achieve similar better": 985,
      "similar better performance": 30099,
      "assess performance llms": 2602,
      "llms present comprehensive": 18862,
      "present comprehensive evaluation": 25521,
      "popular llms llama": 25127,
      "continual learning large": 5976,
      "language models aligned": 16246,
      "models llms demonstrate": 21494,
      "llms demonstrate exceptional": 18509,
      "continual learning benchmarks": 5975,
      "instruction tuning paper": 15204,
      "tuning paper introduce": 34002,
      "benchmark designed evaluate": 3380,
      "empirical findings suggest": 9227,
      "inference acceleration large": 14760,
      "acceleration large language": 805,
      "sparse finetuning large": 30613,
      "llms finetuning pretrained": 18612,
      "finetuning pretrained llms": 11490,
      "methods instruction data": 20052,
      "open source models": 23426,
      "models varying sizes": 22112,
      "language models zeroshot": 16760,
      "time series forecasting": 33142,
      "models llms gpt3": 21545,
      "model size generally": 20791,
      "data collection model": 6643,
      "incontext learning capability": 14551,
      "capability large language": 4094,
      "language models gpt35": 16372,
      "expertise prompt engineering": 10515,
      "conducted user study": 5642,
      "exceptional performance various": 10172,
      "recent studies focused": 27554,
      "llms shedding light": 18935,
      "capability language models": 4092,
      "costs work propose": 6278,
      "human preference datasets": 13855,
      "code dataset model": 4743,
      "billionparameter language model": 3725,
      "language model specialized": 16203,
      "available large language": 2986,
      "language models generation": 16362,
      "multiple dimensions including": 22387,
      "existing language models": 10282,
      "datasets publicly available": 7166,
      "applied question answering": 2197,
      "question answering text": 27058,
      "generation tasks language": 12615,
      "natural language sentences": 22566,
      "improves performance existing": 14388,
      "instruction tuning human": 15192,
      "instruction data generation": 15146,
      "blooms taxonomy classic": 3795,
      "yields significant performance": 35928,
      "benchmarks hope work": 3446,
      "learning process llms": 17635,
      "entity recognition using": 9651,
      "pretrained transformerbased models": 25770,
      "training dataset using": 33496,
      "model llm using": 20626,
      "using dataset train": 34765,
      "based bert model": 3139,
      "fundamental component language": 11977,
      "models llms perform": 21600,
      "perform multiple choice": 24496,
      "choice question answering": 4553,
      "question answering mcqa": 27047,
      "ability llms smaller": 703,
      "llms smaller language": 18956,
      "sentence embedding models": 29533,
      "llms chatgpt llama": 18466,
      "model achieves comparable": 20348,
      "generative ai applications": 12646,
      "stateoftheart large language": 30940,
      "instruction tuning using": 15209,
      "llms like llama": 18764,
      "responses paper propose": 28504,
      "consistently improves performance": 5752,
      "super natural instructions": 31638,
      "teaching language models": 32592,
      "math reasoning tasks": 19673,
      "contrast prior work": 6016,
      "models improve performance": 21373,
      "address limitations present": 1344,
      "limitations present new": 17931,
      "conduct experiments diverse": 5603,
      "experiments diverse set": 10437,
      "diverse set tasks": 8461,
      "zeroshot fewshot scenarios": 35975,
      "supervised learning tasks": 31683,
      "remains poorly understood": 28009,
      "pretrained foundation models": 25645,
      "generating instructiontuning data": 12433,
      "using incontext learning": 34797,
      "wang et al": 35416,
      "models 175b parameters": 20925,
      "proposed method yields": 26610,
      "recent advancements natural": 27496,
      "language processing large": 16784,
      "processing large language": 26108,
      "work provides insights": 35766,
      "provides insights potential": 26758,
      "understanding large language": 34240,
      "models llms remarkable": 21619,
      "recent studies suggest": 27558,
      "gpt35turbo gpt4 llama2": 13044,
      "llama2 series models": 18193,
      "models represent reason": 21929,
      "instruction tuned large": 15184,
      "remains lack comprehensive": 27999,
      "lack comprehensive investigation": 15980,
      "address gap present": 1325,
      "benchmark specifically designed": 3412,
      "evaluate performance various": 9858,
      "multilingual pretrained language": 22324,
      "analysis reveals existing": 1960,
      "instruction tuned llms": 15185,
      "used previous works": 34619,
      "using language models": 34806,
      "method outperforms stateoftheart": 19953,
      "outperforms stateoftheart baselines": 23853,
      "current landscape large": 6497,
      "like llama mistral": 17885,
      "findings shed light": 11254,
      "existing work focuses": 10326,
      "tasks require reasoning": 32483,
      "datasets various settings": 7190,
      "release code pretrained": 27903,
      "code pretrained checkpoints": 4797,
      "open information extraction": 23396,
      "challenging task natural": 4400,
      "methods require significant": 20089,
      "paper introduce novel": 24060,
      "introduce novel framework": 15527,
      "training data furthermore": 33476,
      "reducing training time": 27760,
      "time experimental results": 33124,
      "extensive automatic human": 10680,
      "challenging natural language": 4389,
      "multiple llms including": 22400,
      "llms including vicuna": 18707,
      "retrieval augmented large": 28736,
      "augmented large language": 2820,
      "commercial large language": 4985,
      "incontext learning ability": 14550,
      "evaluate effectiveness proposed": 9832,
      "effectiveness proposed methods": 8966,
      "models language models": 21417,
      "large models finetuning": 17235,
      "training data making": 33483,
      "source domain target": 30558,
      "domain target domains": 8597,
      "language processing computer": 16778,
      "processing computer vision": 26100,
      "vision downstream tasks": 35294,
      "commonly used datasets": 5031,
      "generative modeling tasks": 12676,
      "work bridge gap": 35673,
      "bridge gap proposing": 3867,
      "significantly boosts performance": 30041,
      "standard language modeling": 30879,
      "including autoencoding models": 14457,
      "models encoderdecoder models": 21201,
      "leading improved performance": 17476,
      "compared models like": 5153,
      "achieves superior performance": 1077,
      "different downstream tasks": 8074,
      "benchmarks human evaluation": 3448,
      "evaluation results demonstrate": 10001,
      "information language models": 14876,
      "plays important role": 25029,
      "llama2 70b model": 18161,
      "address problem propose": 1350,
      "scales 7b 13b": 29152,
      "multimodal models multiple": 22364,
      "language models rapid": 16679,
      "models rapid advancement": 21893,
      "rapid advancement large": 27243,
      "advancement large language": 1455,
      "various language models": 35106,
      "synthetic dataset demonstrates": 31854,
      "models llms including": 21562,
      "llms used generate": 19023,
      "feasibility using llms": 11018,
      "potential pretrained large": 25290,
      "llms experimental results": 18587,
      "factors influence performance": 10873,
      "provide insights future": 26710,
      "method large language": 19938,
      "llms shown great": 18939,
      "great potential natural": 13255,
      "conduct comprehensive experiments": 5594,
      "comprehensive experiments demonstrate": 5377,
      "experiments demonstrate effectiveness": 10432,
      "recently released llms": 27620,
      "improving llms performance": 14414,
      "models like gpt4": 21457,
      "believe work provides": 3342,
      "work provides valuable": 35768,
      "valuable insights future": 35013,
      "insights future research": 15075,
      "recent work shown": 27569,
      "convolutional neural network": 6124,
      "exhibit impressive reasoning": 10222,
      "reasoning data augmentation": 27401,
      "tasks small models": 32509,
      "models work propose": 22136,
      "data augmentation ability": 6606,
      "syntactic language models": 31820,
      "classification tasks building": 4615,
      "systems using large": 31925,
      "like llama 7b": 17883,
      "llama 7b 13b": 18067,
      "opensource models achieve": 23529,
      "achieve competitive performance": 955,
      "language processing task": 16799,
      "exhibited remarkable performance": 10237,
      "remarkable performance various": 28048,
      "llms including chatgpt": 18689,
      "including chatgpt gpt4": 14465,
      "leverage user feedback": 17763,
      "quantization large language": 27005,
      "llama2 model family": 18187,
      "enhance model performance": 9520,
      "language models study": 16716,
      "model parameters experiments": 20691,
      "enhance llms ability": 9518,
      "follow user instructions": 11681,
      "performance improvement variety": 24626,
      "average accuracy improvement": 3010,
      "specific downstream tasks": 30692,
      "requiring extensive training": 28271,
      "varying difficulty levels": 35173,
      "resulting significantly improved": 28562,
      "experiments language models": 10455,
      "zeroshot fewshot prompting": 35974,
      "retrieval augmented generation": 28734,
      "augmented generation rag": 2816,
      "learning human preferences": 17581,
      "preference optimization dpo": 25470,
      "pairs preference data": 23983,
      "data demonstrate significant": 6672,
      "challenges future directions": 4347,
      "cost training models": 6256,
      "enlarging model sizes": 9584,
      "model 13 billion": 20333,
      "foundation model pretrained": 11797,
      "significantly outperforms models": 30077,
      "models multiple benchmarks": 21743,
      "reflect differences model": 27790,
      "differences model performance": 8038,
      "encoded large language": 9341,
      "samples large language": 29081,
      "data generation approach": 6723,
      "open large language": 23401,
      "generated synthetic data": 12391,
      "applicability large language": 2121,
      "remains unexplored study": 28022,
      "language models conduct": 16285,
      "different parameter sizes": 8118,
      "model size grows": 20792,
      "despite impressive performance": 7792,
      "substantial computational resources": 31464,
      "computational resources making": 5479,
      "potential address challenges": 25237,
      "address challenges introduce": 1314,
      "orders magnitude larger": 23683,
      "including finetuning incontext": 14478,
      "finetuning incontext learning": 11419,
      "llms demonstrated superior": 18527,
      "llms like gpt3": 18761,
      "significant progress natural": 30011,
      "progress natural language": 26219,
      "stateoftheart llms like": 30949,
      "models various tasks": 22110,
      "demonstrated large language": 7533,
      "language explanations nles": 16072,
      "improve robustness llms": 14296,
      "introduce new approach": 15519,
      "evaluate popular llms": 9860,
      "models llms given": 21542,
      "improved chainofthought prompting": 14308,
      "intermediate computation steps": 15426,
      "fundamental linguistic phenomenon": 11980,
      "models pretrained code": 21840,
      "languages recent advancements": 16911,
      "proliferation large language": 26259,
      "popular large language": 25120,
      "different language families": 8088,
      "languages like english": 16889,
      "generative tasks like": 12707,
      "popularity large language": 25143,
      "recent advances transformerbased": 27506,
      "advances transformerbased large": 1489,
      "transformerbased large language": 33752,
      "great strides natural": 13260,
      "strides natural language": 31149,
      "domains large language": 8624,
      "llms exhibit remarkable": 18581,
      "performance larger models": 24651,
      "proprietary models gpt35": 26648,
      "models gpt35 gpt4": 21330,
      "limitations language model": 17923,
      "language model training": 16208,
      "language model lm": 16173,
      "information training data": 14921,
      "consistently improve performance": 5750,
      "set data samples": 29682,
      "future research direction": 12044,
      "models llms extensive": 21530,
      "language model handle": 16153,
      "diverse contexts different": 8418,
      "source code publicly": 30552,
      "augmented language models": 2818,
      "number parameters language": 23156,
      "language models proven": 16672,
      "models proven effective": 21872,
      "used measure performance": 34612,
      "llms lowresource languages": 18788,
      "downstream tasks unlike": 8701,
      "pretrained word embeddings": 25782,
      "leveraging contextual information": 17781,
      "partofspeech pos tagging": 24367,
      "lm training finetuning": 19064,
      "teaching small language": 32595,
      "language models reason": 16681,
      "outperform conventional instructiontuned": 23771,
      "larger models provide": 17333,
      "help model learn": 13511,
      "advanced reasoning abilities": 1442,
      "llms demonstrated exceptional": 18513,
      "demonstrated exceptional capabilities": 7521,
      "highperformance computing large": 13677,
      "computing large language": 5515,
      "llms including llama": 18702,
      "highperformance computing hpc": 13676,
      "responses response challenge": 28511,
      "response challenge propose": 28474,
      "model supervised finetuning": 20815,
      "generated qa questionanswer": 12380,
      "qa questionanswer instances": 26916,
      "demonstrate comparable performance": 7440,
      "comparable performance existing": 5085,
      "performance existing methods": 24589,
      "bridge performance gap": 3870,
      "utilization language models": 34954,
      "language models model": 16628,
      "result significant performance": 28548,
      "overcome problem propose": 23925,
      "finetuned language model": 11322,
      "pretrained base model": 25629,
      "language models produce": 16666,
      "improve performance text": 14288,
      "questionanswering qa tasks": 27087,
      "qa datasets using": 26909,
      "llm automatically generate": 18273,
      "significantly improve performance": 30057,
      "bleu rouge metrics": 3767,
      "compared model finetuning": 5151,
      "approach finetuning llms": 2284,
      "study introduces novel": 31345,
      "introduces novel approach": 15545,
      "language processing techniques": 16803,
      "demonstrates significantly enhanced": 7571,
      "establishes new stateoftheart": 9777,
      "models paper present": 21795,
      "model sizes ranging": 20800,
      "large langauge models": 16958,
      "subset training data": 31454,
      "facilitate research development": 10844,
      "open language models": 23399,
      "models permissive license": 21813,
      "reasoning abilities large": 27369,
      "previous studies typically": 25886,
      "address issue propose": 1334,
      "provides thorough evaluation": 26765,
      "models conduct extensive": 21089,
      "extensive experiments popular": 10700,
      "gpt4 llama2 mistral": 13092,
      "indicate significant performance": 14696,
      "significant performance gap": 30005,
      "llms specifically designed": 18966,
      "datasets experimental results": 7111,
      "experimental results reveal": 10407,
      "accurate contextually relevant": 922,
      "models llms llms": 21581,
      "multimodal language model": 22350,
      "pretrained visionlanguage model": 25778,
      "reasoning capabilities innovative": 27383,
      "provide comprehensive understanding": 26691,
      "models llms various": 21662,
      "llms various tasks": 19036,
      "highresource languages english": 13708,
      "exhibit superior performance": 10232,
      "comprehensive empirical analysis": 5365,
      "recent advancements generative": 27490,
      "language generation capabilities": 16083,
      "lowresource language use": 19312,
      "instruction dataset covering": 15151,
      "classification question answering": 4607,
      "demonstrating superior performance": 7591,
      "superior performance current": 31651,
      "finetuning llama27b model": 11446,
      "language models approach": 16248,
      "results various tasks": 28709,
      "reducing memory consumption": 27755,
      "survey large language": 31775,
      "llms gpt4 llama": 18656,
      "models 7b 13b": 20929,
      "7b 13b 70b": 528,
      "tackle issue introduce": 31998,
      "introduce novel inference": 15528,
      "novel inference method": 23089,
      "models llms llama": 21580,
      "llms llama falcon": 18769,
      "code technical reports": 4822,
      "code data model": 4731,
      "data model checkpoints": 6771,
      "gain deeper insights": 12057,
      "highlevel concepts represented": 13624,
      "llms consistently outperform": 18487,
      "best opensource models": 3567,
      "50 billion parameters": 420,
      "billion parameters using": 3720,
      "base chat models": 3115,
      "human annotations paper": 13792,
      "proximal policy optimization": 26795,
      "policy optimization ppo": 25088,
      "series opensource llms": 29644,
      "demonstrates exceptional performance": 7559,
      "significantly improves accuracy": 30061,
      "nlp tasks models": 22960,
      "generate meaningful responses": 12302,
      "quantitative qualitative evaluations": 26995,
      "models llms increasingly": 21567,
      "llms increasingly integrated": 18712,
      "ability llms comprehend": 698,
      "llms particularly gpt4": 18846,
      "comparative analysis llms": 5097,
      "llms using human": 19026,
      "providing valuable insights": 26786,
      "models effective text": 21181,
      "complex contextual relationships": 5270,
      "language model meta": 16175,
      "model meta ai": 20640,
      "advancement field natural": 1448,
      "improve natural language": 14281,
      "70 billion parameters": 496,
      "language adaptation strategies": 16039,
      "achieves average improvement": 1030,
      "information large number": 14881,
      "context window models": 5927,
      "limited address issue": 17939,
      "model training data": 20840,
      "training data available": 33469,
      "manually designed prompts": 19571,
      "capabilities current stateoftheart": 4008,
      "knowledge graphs kgs": 15860,
      "policy gradient reinforcement": 25081,
      "gradient reinforcement learning": 13191,
      "reinforcement learning algorithm": 27839,
      "current stateoftheart model": 6535,
      "method code available": 19889,
      "code available github": 4723,
      "language models exploring": 16337,
      "refining large language": 27787,
      "machinegenerated instructionfollowing data": 19370,
      "models exhibit impressive": 21227,
      "paper systematically investigate": 24140,
      "model instruction finetuning": 20588,
      "architecture code data": 2437,
      "data model publicly": 6773,
      "local large language": 19130,
      "superior performance various": 31654,
      "nlp tasks inspired": 22954,
      "available apache 20": 2966,
      "apache 20 license": 2099,
      "language model generate": 16146,
      "learners large language": 17527,
      "llms natural language": 18811,
      "understanding question answering": 34264,
      "finetuned language models": 11323,
      "llms incontext learning": 18709,
      "generative tasks using": 12708,
      "analysis sheds light": 1963,
      "models llms highlights": 21555,
      "llms highlights potential": 18675,
      "language models goal": 16366,
      "scales large language": 29154,
      "prompts extensive experiments": 26419,
      "extensive experiments conducted": 10694,
      "verify effectiveness proposed": 35218,
      "models project page": 21860,
      "project page available": 26246,
      "space recent work": 30582,
      "information learned representations": 14883,
      "representations large language": 28163,
      "representational similarity analysis": 28155,
      "language model architectures": 16120,
      "recent trend large": 27566,
      "scale model size": 29143,
      "stateoftheart language model": 30937,
      "demonstrate proposed approach": 7482,
      "proposed approach significantly": 26591,
      "experiments conducted using": 10428,
      "achieve stateoftheart performance": 991,
      "stateoftheart performance terms": 30975,
      "models paper explores": 21792,
      "tasks sentiment analysis": 32494,
      "different prompting strategies": 8128,
      "results indicate need": 28634,
      "tasks natural language": 32428,
      "30 billion parameters": 307,
      "factual knowledge graph": 10888,
      "code data results": 4740,
      "future research endeavors": 12046,
      "mainstream llms llama": 19411,
      "nonenglish languages paper": 22993,
      "question conduct extensive": 27064,
      "pretraining instruction tuning": 25803,
      "paper presents development": 24098,
      "initial pretraining phase": 14962,
      "field materials science": 11142,
      "available research community": 3000,
      "remarkable capabilities understanding": 28037,
      "opensource language model": 23507,
      "plays crucial role": 25027,
      "suggest continual pretraining": 31568,
      "era generative ai": 9696,
      "hold immense promise": 13735,
      "small language model": 30347,
      "language model present": 16190,
      "significantly outperforms existing": 30076,
      "publicly available github": 26857,
      "end propose new": 9417,
      "knowledge catastrophic forgetting": 15823,
      "performance various benchmarks": 24798,
      "models llama family": 21465,
      "natural programming languages": 22579,
      "laying solid foundation": 17453,
      "development opensource large": 7968,
      "study scaling laws": 31393,
      "advancing opensource language": 1495,
      "conduct supervised finetuning": 5622,
      "sft direct preference": 29761,
      "exhibits superior performance": 10256,
      "outperforms llama 70b": 23830,
      "mathematics code generation": 19688,
      "code generation multilingual": 4762,
      "provide model finetuned": 26715,
      "model finetuned follow": 20525,
      "finetuned follow instructions": 11311,
      "gemini pro llama": 12142,
      "chat model human": 4444,
      "base instruct models": 3118,
      "models released apache": 21919,
      "released apache 20": 27920,
      "gpt35 gpt4 llama2": 13024,
      "spatial reasoning capabilities": 30641,
      "reasoning capabilities llms": 27387,
      "multimodal large language": 22352,
      "models llms multimodal": 21585,
      "llms multimodal large": 18806,
      "language models mllms": 16624,
      "models mllms shown": 21733,
      "tasks address gap": 32237,
      "experimental results models": 10403,
      "language models mental": 16618,
      "models mental health": 21717,
      "mental health challenges": 19840,
      "transformerbased models like": 33764,
      "like bert xlnet": 17851,
      "models llms potential": 21602,
      "using llms generate": 34826,
      "demonstrate impressive capabilities": 7463,
      "diverse downstream tasks": 8427,
      "downstream tasks paper": 8697,
      "mixtureofexperts language models": 20286,
      "language models era": 16322,
      "models era large": 21213,
      "models llms hold": 21557,
      "paper investigates potential": 24074,
      "language models search": 16701,
      "instruction tuning large": 15197,
      "tuning large language": 33988,
      "work explore potential": 35706,
      "empirical results reveal": 9237,
      "models publicly accessible": 21878,
      "question answering training": 27059,
      "frequently asked questions": 11927,
      "reward model train": 28860,
      "using policy gradient": 34874,
      "models based t5": 21009,
      "data natural language": 6781,
      "models human evaluation": 21363,
      "efficient finetuning peft": 9038,
      "finetuning effective way": 11394,
      "instruction tuning datasets": 15190,
      "performance lowresource languages": 24669,
      "paper conduct comprehensive": 24024,
      "conduct comprehensive study": 5596,
      "significantly improved performance": 30059,
      "given training data": 12780,
      "incurs high cost": 14665,
      "llms excel tasks": 18575,
      "address data scarcity": 1321,
      "diverse highquality dataset": 8432,
      "mathematical reasoning capabilities": 19684,
      "wider research community": 35587,
      "retrievalaugmented generation rag": 28762,
      "popular llms including": 25126,
      "llms including llama213b": 18704,
      "questions answers using": 27097,
      "demonstrate finetuned model": 7456,
      "despite general capabilities": 7778,
      "general capabilities large": 12159,
      "language models consistently": 16288,
      "knowledge reasoning safety": 15899,
      "factual knowledge demonstrate": 10887,
      "vision language tasks": 35305,
      "gpt2 models results": 12927,
      "model instruction finetuned": 20587,
      "models trained evaluated": 22068,
      "models human preferences": 21364,
      "training work study": 33645,
      "instruction following ability": 15163,
      "iterations approach yields": 15684,
      "approach yields model": 2365,
      "yields model outperforms": 35926,
      "model outperforms existing": 20672,
      "outperforms existing systems": 23822,
      "text generation capabilities": 32869,
      "generation capabilities llms": 12467,
      "models llms task": 21650,
      "llm training data": 18375,
      "using dataset collected": 34764,
      "llms llama2 mistral": 18778,
      "publicly release code": 26863,
      "release code data": 27902,
      "fusion large language": 12023,
      "models training large": 22079,
      "improve performance target": 14287,
      "code generation code": 4755,
      "model weights data": 20867,
      "weights data public": 35505,
      "complex language tasks": 5278,
      "text generation address": 32867,
      "novel framework designed": 23082,
      "results demonstrate proposed": 28591,
      "bridge gap propose": 3866,
      "propose new task": 26542,
      "benchmark designed assess": 3379,
      "perform new task": 24498,
      "achieves comparable results": 1040,
      "task performance pruning": 32174,
      "roberta t5 models": 28923,
      "trillion tokens sourced": 33908,
      "specific use cases": 30724,
      "associated code publicly": 2644,
      "code publicly accessible": 4802,
      "inspire future research": 15092,
      "future research practical": 12048,
      "research practical applications": 28343,
      "high memory bandwidth": 13575,
      "explainability large language": 10527,
      "models llms critical": 21492,
      "aspect natural language": 2566,
      "language processing llms": 16787,
      "significant concerns regarding": 29973,
      "llms llama family": 18770,
      "pretraining large language": 25811,
      "language models known": 16398,
      "new training procedure": 22862,
      "provide extensive analysis": 26700,
      "aigenerated content aigc": 1634,
      "generate training data": 12336,
      "7b 13b 34b": 527,
      "stateoftheart opensource models": 30968,
      "achieves performance par": 1058,
      "reasoning multimodal large": 27426,
      "large language modelsmllms": 17223,
      "poses significant challenge": 25174,
      "extreme compression large": 10792,
      "compression large language": 5418,
      "size poses significant": 30272,
      "poses significant challenges": 25175,
      "training inference costs": 33532,
      "llama2 7b model": 18164,
      "multilingual capabilities large": 22298,
      "language models models": 16629,
      "extending large language": 10664,
      "data results indicate": 6843,
      "presents promising direction": 25592,
      "encoderdecoder language model": 9367,
      "language model enhanced": 16136,
      "recent advances natural": 27503,
      "advances natural language": 1484,
      "pretrained encoderdecoder architecture": 25641,
      "compute memory resources": 5496,
      "zeroshot task performance": 35998,
      "models code available": 21057,
      "based insights introduce": 3180,
      "dialogue code generation": 8012,
      "finance large language": 11211,
      "using financial domain": 34779,
      "apply supervised finetuning": 2214,
      "13b chat model": 125,
      "augmentation language models": 2801,
      "learning language models": 17591,
      "math reasoning testbed": 19674,
      "significant performance gain": 30003,
      "training curriculum learning": 33465,
      "versatile multimodal large": 35225,
      "language model mllm": 16177,
      "lowrank adaption lora": 19305,
      "compared original lora": 5158,
      "consistent performance gains": 5742,
      "demonstrates significant performance": 7568,
      "nlp tasks propose": 22962,
      "trained supervised finetuning": 33431,
      "used text generation": 34630,
      "generation based gpt2": 12463,
      "aligning large language": 1746,
      "language models news": 16632,
      "biases generated text": 3673,
      "unclear models perform": 34127,
      "research highlights potential": 28324,
      "language models critical": 16292,
      "language models developed": 16309,
      "trillion tokens english": 33907,
      "open language model": 23398,
      "details training data": 7848,
      "including training data": 14523,
      "training data training": 33492,
      "solve wide range": 30500,
      "llms llama2 gpt35": 18775,
      "llama2 gpt35 palm2": 18179,
      "performs par better": 24851,
      "llms 7b 70b": 18403,
      "7b 70b parameters": 531,
      "language models face": 16340,
      "representative large language": 28182,
      "models capable performing": 21043,
      "paper present method": 24089,
      "gpt4 smaller models": 13120,
      "models gpt4 using": 21335,
      "100 success rate": 40,
      "previous methods using": 25870,
      "different sizes gpt2": 8140,
      "holdout test set": 13739,
      "perform extensive experiments": 24489,
      "novel lightweight framework": 23093,
      "llms llama2 gpt4": 18777,
      "models llms particularly": 21599,
      "using blooms taxonomy": 34743,
      "substantial computational memory": 31462,
      "computational memory requirements": 5471,
      "inference recent advancements": 14805,
      "identify current limitations": 14007,
      "current limitations discuss": 6508,
      "future directions improve": 12032,
      "llm inference efficiency": 18322,
      "guardrails large language": 13336,
      "identify mitigate risks": 14013,
      "tasks model sizes": 32420,
      "finetuning peft methods": 11476,
      "model inference sparsityaware": 20581,
      "models demonstrate effectiveness": 21132,
      "downstream tasks experiments": 8692,
      "experiments proposed method": 10468,
      "maintaining competitive performance": 19422,
      "data instruction tuning": 6741,
      "commonsense reasoning reading": 5042,
      "reasoning reading comprehension": 27445,
      "improves performance llama": 14389,
      "code model dataset": 4782,
      "models confidence scores": 21094,
      "benchmarks demonstrate proposed": 3436,
      "single hidden state": 30205,
      "increase number parameters": 14602,
      "minimal computational overhead": 20185,
      "pretraining resulting model": 25835,
      "linear computational complexity": 17988,
      "validate effectiveness approach": 34994,
      "performance multiple benchmarks": 24684,
      "multiple benchmarks code": 22381,
      "model weights datasets": 20869,
      "lottery ticket hypothesis": 19264,
      "work present comprehensive": 35748,
      "models specifically llama2": 22002,
      "model achieves superior": 20352,
      "underscore effectiveness finetuning": 34174,
      "datasets contain short": 7085,
      "benchmark includes datasets": 3391,
      "use model filter": 34551,
      "models llms proven": 21609,
      "llms work propose": 19049,
      "effective training framework": 8904,
      "achieves best performance": 1032,
      "question answering task": 27056,
      "task large language": 32149,
      "task artificial intelligence": 32080,
      "capture contextual information": 4149,
      "directly applying llms": 8234,
      "enhance reasoning abilities": 9529,
      "work propose alternative": 35752,
      "sparsity large language": 30631,
      "natural approach reduce": 22504,
      "approach reduce cost": 2333,
      "inference existing methods": 14776,
      "like gpt llama": 17866,
      "strong performance wide": 31186,
      "massive amounts text": 19624,
      "llms including popular": 18706,
      "compare performance popular": 5112,
      "performance popular llms": 24716,
      "open challenges future": 23387,
      "challenges future research": 4348,
      "models diverse set": 21165,
      "instructions instruction finetuning": 15254,
      "instruction finetuning ift": 15158,
      "framework future research": 11859,
      "knowledge graph completion": 15858,
      "language model recent": 16195,
      "language models achieved": 16238,
      "achieved stateoftheart performance": 1018,
      "stateoftheart performance multiple": 30972,
      "llms reasoning abilities": 18891,
      "enhance performance llms": 9525,
      "llms performance various": 18851,
      "reasoning performance llms": 27434,
      "models including gpt35": 21382,
      "training language model": 33538,
      "llama model significantly": 18130,
      "model significantly outperforms": 20784,
      "llms generative ai": 18640,
      "models llms great": 21552,
      "llms great potential": 18664,
      "different llms gpt4": 8097,
      "received lot attention": 27482,
      "languages lowresource languages": 16894,
      "improve language model": 14271,
      "language model performance": 16187,
      "model finetuned model": 20532,
      "finetuned model shows": 11344,
      "llms increasingly utilized": 18714,
      "models previous studies": 21851,
      "llama2 falcon mistral": 18174,
      "finetuned models exhibit": 11346,
      "language models abstractive": 16236,
      "applied large language": 2192,
      "additionally qualitative analysis": 1300,
      "extensive experiments multiple": 10699,
      "using gpt 35": 34788,
      "demonstrate significant performance": 7494,
      "downstream tasks given": 8694,
      "models enabling use": 21197,
      "gpu memory requirements": 13176,
      "experiments llama2 mistral": 10460,
      "understanding incontext learning": 34233,
      "language processing based": 16776,
      "latent variable models": 17412,
      "common sense knowledge": 5013,
      "nlp tasks empirical": 22950,
      "understanding language models": 34238,
      "nexttoken probabilities computed": 22897,
      "precision recall assess": 25394,
      "llms paper introduces": 18840,
      "paper introduces novel": 24064,
      "introduces novel evaluation": 15546,
      "novel evaluation framework": 23078,
      "evaluation framework large": 9949,
      "framework large language": 11875,
      "quality diversity generated": 26955,
      "comprehensive evaluation stateoftheart": 5374,
      "models finetuned human": 21276,
      "poorly understood paper": 25109,
      "question answering cqa": 27041,
      "llms gpt 35": 18644,
      "gpt 35 llama": 12844,
      "paper addresses challenge": 24005,
      "demonstrating significant improvement": 7588,
      "introduce novel approach": 15525,
      "social media posts": 30426,
      "using open source": 34864,
      "open source large": 23423,
      "source large language": 30565,
      "language model llama2": 16164,
      "power natural language": 25327,
      "research focuses developing": 28319,
      "language model provides": 16194,
      "context address challenge": 5881,
      "address challenge propose": 1312,
      "prompts large language": 26427,
      "model performance notably": 20703,
      "additionally findings reveal": 1286,
      "challenges large language": 4355,
      "scenarios address gap": 29202,
      "address gap introduce": 1324,
      "llms like gpt35": 18762,
      "contexts language models": 5943,
      "models lms strong": 21680,
      "reasoning knowledge graph": 27414,
      "paper aim improve": 24009,
      "improve reasoning ability": 14293,
      "models llms knowledge": 21570,
      "llms knowledge graphs": 18736,
      "datasets code data": 7075,
      "data publicly released": 6822,
      "models llms new": 21592,
      "challenge paper propose": 4325,
      "new evaluation benchmark": 22800,
      "set evaluation metrics": 29684,
      "experimental evaluation shows": 10385,
      "models dont learn": 21171,
      "present empirical investigation": 25527,
      "model family llama": 20516,
      "models llms make": 21582,
      "introduce new task": 15522,
      "benchmark comprehensive evaluation": 3360,
      "comprehensive evaluation benchmark": 5368,
      "perform better tasks": 24473,
      "graph neural network": 13226,
      "network large language": 22695,
      "learning icl capabilities": 17585,
      "neural network gnn": 22744,
      "classification tasks gpt2": 4616,
      "llms demonstrated strong": 18526,
      "capable llms like": 4112,
      "outperform strong baselines": 23789,
      "language models gpt4": 16373,
      "languages large language": 16885,
      "crosslingual knowledge transfer": 6415,
      "evaluate different llms": 9826,
      "comprehension generation tasks": 5343,
      "enhance multilingual capabilities": 9523,
      "finetuned llms using": 11340,
      "using lowrank adaptation": 34831,
      "models natural language": 21745,
      "new approach generating": 22777,
      "models trained data": 22066,
      "model trained data": 20835,
      "indomain training data": 14727,
      "usage large language": 34506,
      "recently emerged promising": 27595,
      "models lms proven": 21678,
      "machine translation paper": 19363,
      "llms pretrained large": 18865,
      "massive multitask language": 19628,
      "pretraining large models": 25813,
      "efficiency large language": 9006,
      "recent efforts explored": 27517,
      "help llms achieve": 13508,
      "comparable model performance": 5083,
      "model performance paper": 20704,
      "llama27b llama213b respectively": 18207,
      "achieving superior performance": 1110,
      "stochastic beam search": 31070,
      "computational cost llm": 5459,
      "llms demonstrate remarkable": 18511,
      "potential various domains": 25308,
      "highquality instruction data": 13691,
      "bilingual english chinese": 3709,
      "existing datasets introduce": 10268,
      "performance llms especially": 24664,
      "language models external": 16338,
      "machine learning model": 19348,
      "finetuned llama27b model": 11337,
      "llms exhibit different": 18579,
      "study large language": 31353,
      "multiple choice questions": 22384,
      "cognitive abilities knowledge": 4875,
      "model size paper": 20794,
      "textual data augmentation": 33025,
      "data augmentation da": 6607,
      "tasks paper challenge": 32441,
      "address limitation propose": 1342,
      "propose simple approach": 26566,
      "tokens encode information": 33223,
      "model achieve stateoftheart": 20344,
      "recently proposed address": 27615,
      "proposed address issue": 26589,
      "exhibits significant performance": 10251,
      "significant performance drops": 30002,
      "compared standard finetuning": 5171,
      "parameters propose simple": 24282,
      "significant performance gains": 30004,
      "single a100 gpu": 30197,
      "tokens large language": 33238,
      "remarkable capabilities various": 28038,
      "decoding process address": 7281,
      "proposed framework significantly": 26601,
      "significant advancement field": 29951,
      "remarkable capabilities language": 28032,
      "mistral zephyr models": 20236,
      "performance levels comparable": 24656,
      "finetuned models findings": 11347,
      "findings underscore potential": 11260,
      "valuable resource understanding": 35017,
      "llama vicuna mistral": 18152,
      "models heavily relies": 21350,
      "highquality pretraining data": 13698,
      "improve data quality": 14261,
      "example use cases": 10113,
      "improving data quality": 14406,
      "evaluate large language": 9842,
      "tasks limited understanding": 32405,
      "understanding llms perform": 34247,
      "intellectual property ip": 15349,
      "language model called": 16124,
      "benchmark experimental results": 3387,
      "code data models": 4734,
      "data models available": 6777,
      "small subset neurons": 30370,
      "language models attention": 16250,
      "models accurately predict": 20943,
      "language model representations": 16197,
      "substantial computational costs": 31461,
      "novel approach designed": 23057,
      "approach designed reduce": 2258,
      "reduce computational costs": 27704,
      "language model series": 16200,
      "models available hugging": 21002,
      "language models explore": 16335,
      "models explore approach": 21241,
      "models plms bert": 21820,
      "series flant5 llama": 29637,
      "models llms ability": 21473,
      "time large language": 33131,
      "single forward pass": 30202,
      "role attention heads": 28954,
      "llms text generation": 19003,
      "text generation large": 32872,
      "models llms known": 21571,
      "paper propose simple": 24116,
      "models proposed framework": 21870,
      "knowledge distillation additional": 15831,
      "method reinforcement learning": 19967,
      "develop new evaluation": 7919,
      "new evaluation dataset": 22801,
      "propose novel evaluation": 26547,
      "demonstrates strong performance": 7573,
      "llms code data": 18470,
      "directly prompting llms": 8243,
      "lack domainspecific knowledge": 15987,
      "experiment results demonstrate": 10380,
      "feedback reinforcement learning": 11070,
      "rapidly evolving field": 27256,
      "presents formidable challenge": 25586,
      "gpt35 gpt4 llama27b": 13025,
      "capabilities smaller models": 4071,
      "compared larger counterparts": 5144,
      "natural language input": 22527,
      "models achieved remarkable": 20950,
      "pretraining instruction finetuning": 25802,
      "language adaptation large": 16036,
      "adaptation large language": 1181,
      "model foundation model": 20537,
      "empirical results analysis": 9233,
      "resources publicly available": 28445,
      "play crucial role": 25018,
      "construct instruction tuning": 5799,
      "release dataset model": 27907,
      "generalization incontext learning": 12217,
      "paper try answer": 24144,
      "try answer question": 33947,
      "tasks maintaining comparable": 32413,
      "maintaining comparable performance": 19420,
      "models recent advances": 21905,
      "llms reasoning capabilities": 18892,
      "finetuning single gpu": 11527,
      "paper presents new": 24102,
      "results extensive experiments": 28612,
      "hallucination code data": 13374,
      "data evaluation benchmark": 6698,
      "language models minimal": 16622,
      "models minimal human": 21724,
      "minimal human effort": 20187,
      "serving large language": 29665,
      "issue particularly pronounced": 15659,
      "empirical results suggest": 9238,
      "llama 13b model": 18060,
      "paper investigate basic": 24068,
      "models recent works": 21907,
      "models work study": 22137,
      "bias gradient descent": 3647,
      "different prompting techniques": 8129,
      "chain thought cot": 4291,
      "models like gpt35": 21455,
      "tasks extensive experiments": 32326,
      "future large language": 12036,
      "language models surprisingly": 16721,
      "llms perform task": 18848,
      "research question paper": 28355,
      "paper proposes new": 24124,
      "grammatical error correction": 13207,
      "outperform previous stateoftheart": 23783,
      "error correction models": 9712,
      "keyvalue kv cache": 15795,
      "llm inference engine": 18323,
      "language models automatically": 16255,
      "llms transformerbased models": 19009,
      "popular llms chatgpt": 25124,
      "despite considerable advancements": 7773,
      "importance data quality": 14187,
      "data quality quantity": 6824,
      "data synthetic data": 6888,
      "data diverse sources": 6682,
      "language models potentially": 16658,
      "models potentially used": 21828,
      "study aimed develop": 31298,
      "evaluate performance model": 9856,
      "evaluating language models": 9898,
      "question answering code": 27040,
      "code reproduce experiments": 4812,
      "transformerbased language model": 33748,
      "deep learning methods": 7331,
      "zeroshot performance new": 35988,
      "models llms stand": 21644,
      "era artificial intelligence": 9694,
      "computational cost paper": 5460,
      "language models key": 16396,
      "competitive performance stateoftheart": 5226,
      "compared 350m parameter": 5119,
      "350m parameter opt": 346,
      "code available soon": 4724,
      "previous state art": 25880,
      "models demonstrate strong": 21133,
      "demonstrate strong performance": 7501,
      "play vital role": 25021,
      "instructions reinforcement learning": 15270,
      "feedback rlhf framework": 11072,
      "instruction data training": 15149,
      "paving way single": 24423,
      "compared strong baselines": 5174,
      "opened new opportunities": 23455,
      "rouge bleu meteor": 28979,
      "llama2 language models": 18182,
      "based cosine similarity": 3147,
      "tasks text generation": 32528,
      "llms gpt llama": 18645,
      "models llms specifically": 21643,
      "models case study": 21047,
      "models llms use": 21656,
      "experimental results showed": 10408,
      "models opt bloom": 21775,
      "models perform better": 21808,
      "addressing gap introduce": 1372,
      "finetuning llama2 models": 11444,
      "model code data": 20423,
      "exhibit different levels": 10214,
      "paper establish benchmark": 24039,
      "results popular llms": 28657,
      "data curation pipeline": 6667,
      "dataset trained model": 7048,
      "capabilities llm experiments": 4039,
      "like gpt35 llama2": 17872,
      "rapid advancement generative": 27240,
      "advancement generative artificial": 1452,
      "artificial intelligence genai": 2537,
      "high performance computing": 13577,
      "model llm inference": 20623,
      "guide autoregressive generation": 13344,
      "causal language model": 4241,
      "issues propose data": 15673,
      "model shows significant": 20782,
      "robust generalization ability": 28934,
      "generalization ability different": 12206,
      "downstream tasks requires": 8699,
      "instructionfinetuned large language": 15219,
      "downstream task performance": 8686,
      "aware instruction tuning": 3043,
      "especially lowresource languages": 9744,
      "compared competitive baseline": 5127,
      "general task performance": 12186,
      "code models released": 4792,
      "foundation models recently": 11800,
      "agents significantly outperform": 1575,
      "emergence numerous large": 9176,
      "numerous large language": 23185,
      "processing nlp applications": 26116,
      "models finetuning llms": 21283,
      "properties large language": 26475,
      "small medium large": 30355,
      "increase model size": 14600,
      "models significantly better": 21975,
      "counter speech generation": 6292,
      "training downstream tasks": 33504,
      "improves downstream task": 14373,
      "llm generate synthetic": 18314,
      "significantly enhances performance": 30052,
      "student model large": 31255,
      "contemporary large language": 5843,
      "performance existing llms": 24588,
      "language models achieve": 16237,
      "achieve best performance": 946,
      "encoders like bert": 9379,
      "use natural language": 34555,
      "language models billions": 16265,
      "models billions parameters": 21029,
      "zeroshot performance various": 35989,
      "compared prior work": 5166,
      "comprehensive comparison multiple": 5361,
      "incontext learning gpt35": 14554,
      "ablation study demonstrates": 738,
      "benchmarks including truthfulqa": 3450,
      "widespread adoption large": 35592,
      "adoption large language": 1409,
      "models llms facilitated": 21531,
      "generation rag emerged": 12588,
      "users information needs": 34693,
      "retrieval paper propose": 28751,
      "models llms understanding": 21655,
      "various llms including": 35114,
      "including gpt4 llama2": 14492,
      "addition propose new": 1246,
      "artificial intelligence gai": 2536,
      "language models factual": 16341,
      "evaluated various language": 9886,
      "using pretrained language models": 34880,
      "pretrained language models lms": 25679,
      "language models lms various": 16611,
      "models lms various natural": 21684,
      "lms various natural language": 19124,
      "various natural language processing": 35126,
      "natural language processing tasks": 22556,
      "neural machine translation nmt": 22733,
      "language models large language": 16405,
      "models large language models": 21422,
      "masked language models mlms": 19617,
      "largescale pretrained language models": 17376,
      "pretrained language models demonstrated": 25669,
      "language models demonstrated impressive": 16301,
      "models achieved stateoftheart results": 20953,
      "models large pretrained language": 21427,
      "large pretrained language models": 17261,
      "natural language understanding tasks": 22575,
      "neural language models trained": 22729,
      "generative pretrained language model": 12687,
      "pretrained language model gpt2": 25659,
      "pretrained language models paper": 25682,
      "language models paper presents": 16645,
      "paper presents empirical study": 24100,
      "pretrained language models plms": 25685,
      "texttotext transfer transformer t5": 33018,
      "language models lms bert": 16598,
      "neural network language models": 22746,
      "propose new method called": 26540,
      "generative pretrained transformer gpt2": 12695,
      "natural language processing nlp": 22547,
      "recurrent neural networks rnns": 27685,
      "bidirectional encoder representations transformers": 3694,
      "encoder representations transformers bert": 9356,
      "generative language models gpt2": 12665,
      "largescale language models lms": 17362,
      "language models lms able": 16597,
      "advanced neural language models": 1439,
      "pretrained neural language models": 25740,
      "language models lms prone": 16605,
      "gpt2 radford et al": 12944,
      "radford et al 2019": 27152,
      "language models lms demonstrated": 16599,
      "models lms demonstrated impressive": 21672,
      "language model like gpt2": 16163,
      "range natural language understanding": 27204,
      "natural language understanding nlu": 22572,
      "language models bert xlnet": 16261,
      "extremely large batch sizes": 10800,
      "pretrained language models perform": 25683,
      "natural language generation tasks": 22524,
      "parameters constant computational cost": 24236,
      "neural language model gpt2": 22726,
      "performance natural language processing": 24689,
      "use pretrained language models": 34562,
      "language models lms t5": 16609,
      "pretrained language models recent": 25690,
      "language models recent years": 16689,
      "size pretrained language models": 30276,
      "downstream tasks experimental results": 8691,
      "pretraining language understanding generation": 25809,
      "stateoftheart results various natural": 30988,
      "results various natural language": 28707,
      "language processing nlp tasks": 16796,
      "natural language understanding generation": 22570,
      "language understanding generation tasks": 16843,
      "based large language model": 3187,
      "tasks conduct extensive experiments": 32274,
      "transformerbased pretrained language models": 33767,
      "language models lms exhibit": 16601,
      "named entity recognition ner": 22487,
      "significant progress recent years": 30015,
      "datasets demonstrate superior performance": 7093,
      "language processing nlp domain": 16792,
      "general language understanding evaluation": 12175,
      "language models pretrained language": 16662,
      "models pretrained language models": 21842,
      "wide range natural language": 35558,
      "range natural language processing": 27201,
      "adapting pretrained language models": 1217,
      "recent years pretrained language": 27580,
      "years pretrained language models": 35897,
      "pretrained language models bert": 25665,
      "language models bert gpt2": 16259,
      "bert roberta gpt2 dozens": 3530,
      "roberta gpt2 dozens datasets": 28919,
      "current pretrained language models": 6526,
      "language models generate highquality": 16359,
      "models generate highquality text": 21309,
      "language models catastrophic forgetting": 16275,
      "language models bert t5": 16260,
      "language models large pretrained": 16408,
      "outperforms models comparable size": 23836,
      "training large language models": 33545,
      "large language models new": 17177,
      "make code models publicly": 19456,
      "code models publicly available": 4791,
      "adaptation pretrained language models": 1189,
      "remarkable success large language": 28058,
      "success large language models": 31516,
      "frozen pretrained language model": 11941,
      "inference latency experimental results": 14788,
      "large language models llms": 17057,
      "natural language processing models": 22546,
      "machine learning models large": 19350,
      "learning models large language": 17610,
      "large language models llm": 17055,
      "finetuning pretrained language models": 11489,
      "pretrained language models recently": 25692,
      "natural language generation nlg": 22523,
      "data source code available": 6872,
      "structures neural language models": 31234,
      "recurrent neural network rnn": 27683,
      "chen et al 2021": 4531,
      "language model outperforms gpt2": 16182,
      "et al 2019 gpt3": 9793,
      "al 2019 gpt3 brown": 1684,
      "2019 gpt3 brown et": 225,
      "gpt3 brown et al": 12985,
      "brown et al 2020": 3899,
      "generalization natural language processing": 12222,
      "language processing nlp algorithms": 16790,
      "transformerbased language models lms": 33750,
      "language models lms shown": 16607,
      "language models paper introduces": 16643,
      "despite order magnitude smaller": 7800,
      "automated natural language generation": 2872,
      "natural language generation metrics": 22521,
      "large language models present": 17186,
      "berts masked language modeling": 3548,
      "masked language modeling mlm": 19613,
      "prompting large language model": 26383,
      "large language model llm": 16972,
      "question answering natural language": 27049,
      "answering natural language inference": 2067,
      "challenge natural language processing": 4321,
      "language processing nlp systems": 16795,
      "translation summarization question answering": 33852,
      "conduct comprehensive empirical study": 5592,
      "achieve new stateoftheart performance": 975,
      "natural language inference nli": 22526,
      "demonstrate large language models": 7468,
      "large language models chatgpt": 16998,
      "field natural language processing": 11145,
      "glancing language model glm": 12789,
      "generalpurpose pretrained language models": 12259,
      "pretrained language models gpt2": 25673,
      "language models bert albert": 16258,
      "remarkable success natural language": 28061,
      "pretrained transformerbased language models": 25769,
      "language models widely used": 16757,
      "language understanding nlu natural": 16845,
      "understanding nlu natural language": 34256,
      "nlu natural language generation": 22973,
      "batch size learning rate": 3289,
      "autoregressive language models gpt2": 2946,
      "tasks summarization machine translation": 32521,
      "powered large language models": 25334,
      "current stateoftheart sota models": 6538,
      "generative pretrained language models": 12689,
      "model achieves stateoftheart performance": 20351,
      "benefit using large language": 3486,
      "using large language models": 34810,
      "llms 100 billion parameters": 18400,
      "recently large language models": 27608,
      "language understanding nlu tasks": 16847,
      "transformers shown remarkable success": 33797,
      "models generative pretrained transformer": 21317,
      "generative pretrained transformer gpt": 12693,
      "high bandwidth memory hbm": 13554,
      "recent large language models": 27525,
      "language models llms demonstrated": 16446,
      "models llms demonstrated remarkable": 21504,
      "larger language models llms": 17323,
      "large language models large": 17049,
      "language models llms shown": 16561,
      "variety natural language tasks": 35067,
      "improve model performance generalization": 14277,
      "model performance generalization unseen": 20701,
      "performance generalization unseen tasks": 24613,
      "evaluation large language models": 9967,
      "large language models understand": 17215,
      "language models 13b parameters": 16233,
      "large language models increasingly": 17046,
      "multiple natural language tasks": 22405,
      "large language models lms": 17169,
      "models long short term": 21689,
      "long short term memory": 19181,
      "pretrained causal language models": 25635,
      "language models better understand": 16263,
      "large neural language models": 17248,
      "openaccess multilingual language model": 23431,
      "language model large language": 16160,
      "model large language models": 20604,
      "language models llms led": 16510,
      "shown great performance tasks": 29881,
      "performance various nlp tasks": 24807,
      "adapting large language models": 1210,
      "finetuning large language models": 11431,
      "large language models different": 17012,
      "leverage large language models": 17754,
      "large language models diverse": 17013,
      "landscape large language models": 16025,
      "widelyused pretrained language models": 35582,
      "knowledge large language models": 15873,
      "language models llms trained": 16580,
      "using masked language modeling": 34838,
      "masked language modeling task": 19614,
      "knowledge generative language models": 15856,
      "largescale generative language models": 17355,
      "generative language models glms": 12664,
      "language models shown perform": 16703,
      "tasks named entity recognition": 32425,
      "enables pretrained language models": 9308,
      "approaches rely vast amounts": 2390,
      "language models llms surprisingly": 16576,
      "safety large language models": 29050,
      "large language models work": 17220,
      "pretrained texttotext language models": 25755,
      "lack highquality training data": 15992,
      "code data publicly available": 4738,
      "datasets experiment results proposed": 7109,
      "pretrained language generation models": 25655,
      "effectiveness large language models": 8952,
      "performance various natural language": 24803,
      "tasks question answering summarization": 32470,
      "summarization large language models": 31616,
      "language models llms used": 16586,
      "generation large language models": 12535,
      "pretrained language models like": 25676,
      "language models like bert": 16417,
      "like bert gpt t5": 17847,
      "size large language models": 30257,
      "release models research community": 27915,
      "pretrained language models plm": 25684,
      "language models trained large": 16733,
      "use large language models": 34544,
      "language models llms chatgpt": 16437,
      "scaling large language models": 29169,
      "large language models empirical": 17016,
      "language models empirical study": 16319,
      "significantly enhances models performance": 30051,
      "performance large language models": 24646,
      "large language models based": 16995,
      "data large language models": 6751,
      "language models llms downstream": 16457,
      "language models llms revolutionized": 16555,
      "models llms revolutionized natural": 21624,
      "llms revolutionized natural language": 18925,
      "revolutionized natural language processing": 28850,
      "automated machine learning automl": 2867,
      "language models llms gpt4": 16488,
      "large language models success": 17206,
      "language models llms like": 16512,
      "models llms like gpt4": 21578,
      "tasks like image captioning": 32401,
      "feedback large language models": 11066,
      "models llms like chatgpt": 21575,
      "large language models efficient": 17015,
      "potential large language model": 25268,
      "generated large language models": 12369,
      "make code publicly available": 19459,
      "make model data code": 19474,
      "model data code publicly": 20450,
      "data code publicly available": 6637,
      "instruction following large language": 15167,
      "following large language model": 11697,
      "large language models crucial": 17004,
      "research field natural language": 28314,
      "artificial general intelligence agi": 2532,
      "research large language models": 28332,
      "large language models llama": 17054,
      "applications large language models": 2161,
      "datasets large language models": 7140,
      "language models training data": 16735,
      "deploying large language models": 7641,
      "language models llms challenging": 16436,
      "data achieve comparable performance": 6590,
      "ability large language models": 691,
      "era large language models": 9699,
      "pretrained language models ptlms": 25689,
      "descriptions large language models": 7690,
      "large language models temporal": 17208,
      "language models llms multiple": 16524,
      "training data compared baseline": 33472,
      "graphical user interfaces guis": 13236,
      "large language models follow": 17026,
      "language models follow instructions": 16352,
      "leveraging large language models": 17785,
      "prediction large language models": 25428,
      "language models llms produce": 16543,
      "enhancing large language models": 9564,
      "advancements large language models": 1467,
      "large language models given": 17037,
      "alignment large language models": 1767,
      "large language models trained": 17211,
      "end tasks user preferences": 9421,
      "harnessing capabilities large language": 13461,
      "capabilities large language models": 4034,
      "pretrained models bert gpt2": 25720,
      "generative large language models": 12669,
      "models llms revolutionized field": 21623,
      "comes significant computational costs": 4973,
      "significant computational costs paper": 29969,
      "proprietary large language models": 26643,
      "evaluating large language models": 9902,
      "large language models spoken": 17204,
      "models spoken language understanding": 22006,
      "language models demonstrated strong": 16302,
      "spoken language understanding slu": 30817,
      "language models llms directly": 16456,
      "text generated large language": 32863,
      "language models llms data": 16442,
      "furthermore conduct human evaluation": 11991,
      "automatic speech recognition asr": 2896,
      "bias large language models": 3652,
      "comprehensive evaluation large language": 5371,
      "large language models automatic": 16994,
      "make data code publicly": 19463,
      "tasks large language models": 32392,
      "tasks like question answering": 32403,
      "language models lms powerful": 16604,
      "large language models recent": 17194,
      "language models recent advancements": 16684,
      "models recent advancements large": 21903,
      "recent advancements large language": 27493,
      "language models llms significantly": 16569,
      "learning large language models": 17594,
      "pretrained multilingual language models": 25734,
      "large language models partially": 17183,
      "language models llms acquire": 16432,
      "capabilities pretrained language models": 4063,
      "pretrained large language models": 25700,
      "performance variety language tasks": 24795,
      "excel various natural language": 10157,
      "incontext learning instruction tuning": 14561,
      "pretrained language models nlp": 25680,
      "especially large language models": 9741,
      "large language models gpt3": 17039,
      "large language models paper": 17180,
      "language models llms study": 16575,
      "models large language modelsllms": 21425,
      "entities pretrained language models": 9637,
      "using large language model": 34809,
      "multilingual large language models": 22315,
      "language models llms achieved": 16428,
      "models llms achieved great": 21476,
      "llms achieved great success": 18414,
      "evaluating large language model": 9901,
      "language model llm based": 16166,
      "ability generative language models": 680,
      "translation large language models": 33832,
      "recent years large language": 27577,
      "years large language models": 35893,
      "large language models open": 17179,
      "gpt4 metas llama googles": 13097,
      "analysis large language models": 1936,
      "large language models general": 17031,
      "large multilingual language models": 17244,
      "language large language models": 16107,
      "generative artificial intelligence ai": 12655,
      "emergence large language models": 9171,
      "capabilities natural language processing": 4049,
      "generative pretrained transformers gpts": 12700,
      "potential artificial general intelligence": 25242,
      "data augmentation method generate": 6610,
      "pretrained language model plm": 25660,
      "language model plm t5": 16189,
      "new large language model": 22815,
      "large language model code": 16964,
      "process reduces computational requirements": 26081,
      "empowering large language models": 9271,
      "llms including gpt4 llama": 18700,
      "finetuning parameterefficient finetuning peft": 11472,
      "instructiontuned large language model": 15288,
      "large language model based": 16962,
      "language model based llama": 16123,
      "generative pretrained transformer gpt4": 12697,
      "code generation machine translation": 4760,
      "language models llms capture": 16435,
      "llms demonstrated remarkable potential": 18524,
      "experimental results demonstrate superior": 10397,
      "results demonstrate superior performance": 28595,
      "outperforms previous stateoftheart models": 23846,
      "nlp tasks large language": 22958,
      "language models llms typically": 16583,
      "large language models generating": 17034,
      "abilities large language models": 630,
      "large language models providing": 17190,
      "skills large language models": 30312,
      "large language models current": 17005,
      "developments large language models": 7982,
      "language models llms enabled": 16461,
      "various natural language tasks": 35129,
      "language models llms recently": 16546,
      "models llms recently achieved": 21612,
      "following natural language instructions": 11700,
      "language models llms widely": 16592,
      "generating fluent coherent text": 12425,
      "pretrained large language model": 25699,
      "models llms shown remarkable": 21635,
      "effectiveness various generaldomain natural": 8973,
      "various generaldomain natural language": 35098,
      "generaldomain natural language processing": 12193,
      "instructiontuned large language models": 15290,
      "large language models multiple": 17176,
      "reinforcement learning human feedback": 27842,
      "development large language models": 7959,
      "supervised finetuning sft reinforcement": 31680,
      "finetuning sft reinforcement learning": 11522,
      "sft reinforcement learning human": 29766,
      "learning human feedback rlhf": 17579,
      "achieves new stateoftheart result": 1054,
      "reasoning large language models": 27417,
      "language models llms exhibit": 16467,
      "opensource large language model": 23511,
      "great success large language": 13264,
      "large language models computer": 17001,
      "current large language models": 6504,
      "recent advances large language": 27501,
      "advances large language models": 1480,
      "existing large language models": 10285,
      "instructionfollowing large language models": 15232,
      "models pretrained large language": 21844,
      "large language models gpt": 17038,
      "language model gpt2 sequence": 16151,
      "behavior large language models": 3317,
      "supervised finetuning reinforcement learning": 31677,
      "finetuning reinforcement learning human": 11505,
      "machine learning ml models": 19347,
      "diverse natural language processing": 8443,
      "like bert roberta gpt2": 17850,
      "advanced natural language processing": 1435,
      "language processing nlp research": 16794,
      "instructions large language models": 15258,
      "language models llms present": 16539,
      "large language model evaluation": 16965,
      "data codes publicly available": 6640,
      "advent large language models": 1511,
      "revolutionized field natural language": 28846,
      "natural language processing enabling": 22540,
      "large language models focus": 17025,
      "significant advancements natural language": 29956,
      "advancements natural language processing": 1472,
      "models range natural language": 21887,
      "reinforcement learning rl based": 27846,
      "intelligence large language models": 15360,
      "large language models language": 17048,
      "recent progress large language": 27543,
      "progress large language models": 26214,
      "artificial intelligence ai based": 2535,
      "chainofthought cot think stepbystep": 4300,
      "vast amounts text data": 35183,
      "understanding generating humanlike text": 34226,
      "technical report large language": 32609,
      "report large language models": 28119,
      "models llms achieved remarkable": 21478,
      "llms achieved remarkable success": 18419,
      "large language models despite": 17008,
      "language models despite impressive": 16305,
      "despite impressive capabilities large": 7790,
      "impressive capabilities large language": 14233,
      "models llms gpt4 palm": 21550,
      "llms gpt4 palm llama": 18661,
      "fake news detection using": 10936,
      "finetuned large language model": 11326,
      "large language model paper": 16984,
      "model paper considers possibility": 20683,
      "finetuning llama large language": 11442,
      "finetuning peftlora based approach": 11480,
      "peftlora based approach used": 24442,
      "based approach used study": 3135,
      "approach used study model": 2353,
      "used study model finetuned": 34626,
      "study model finetuned following": 31364,
      "model finetuned following tasks": 20528,
      "finetuned following tasks analysing": 11314,
      "following tasks analysing text": 11706,
      "sentiments obtained results finetuned": 29580,
      "obtained results finetuned llama": 23259,
      "results finetuned llama model": 28615,
      "finetuned llama model perform": 11331,
      "extracted sentiments named entities": 10754,
      "sentiments named entities considered": 29576,
      "named entities considered predictive": 22483,
      "entities considered predictive features": 9632,
      "considered predictive features supervised": 5720,
      "predictive features supervised machine": 25451,
      "features supervised machine learning": 11041,
      "supervised machine learning models": 31686,
      "family large language models": 10977,
      "smaller transformerbased language models": 30401,
      "language models llms generate": 16480,
      "llms including gpt35 gpt4": 18695,
      "gpt2 pretrained language model": 12939,
      "pretrained language model corpus": 25658,
      "astronomy large language models": 2674,
      "exploring large language models": 10620,
      "language models llms gpt": 16484,
      "language models llms need": 16526,
      "language models continue advance": 16291,
      "evaluate capabilities language models": 9822,
      "address gap propose novel": 1327,
      "language models llms excel": 16464,
      "models llms excel various": 21524,
      "paper propose novel framework": 24115,
      "closed opensource llms including": 4681,
      "foundational large language models": 11807,
      "power large language models": 25322,
      "large language models dynamic": 17014,
      "inference large language models": 14783,
      "making large language models": 19509,
      "large language models demonstrate": 17006,
      "llms demonstrated remarkable performance": 18523,
      "opensource models similar size": 23535,
      "large language models exhibit": 17018,
      "questions large language models": 27119,
      "experimental results demonstrate effectiveness": 10393,
      "enhance capabilities large language": 9507,
      "large language models using": 17216,
      "large language models research": 17199,
      "impact large language models": 14129,
      "machine translation large language": 19357,
      "machine translation recent work": 19365,
      "machine translation nmt systems": 19362,
      "cases large language models": 4202,
      "language models llms emerged": 16459,
      "language models llms represent": 16553,
      "best performing models achieved": 3572,
      "performing models achieved accuracy": 24837,
      "calculations large language models": 3970,
      "autoregressive large language models": 2949,
      "text classification sentiment analysis": 32827,
      "reasoning ability llms large": 27375,
      "ability llms large language": 701,
      "llms large language models": 18745,
      "remarkable performance wide range": 28050,
      "pose challenges practical deployment": 25160,
      "smaller models experimental results": 30387,
      "language models llms gained": 16477,
      "small language models slms": 30350,
      "large language models solving": 17202,
      "models llms shown promise": 21634,
      "prompting large language models": 26384,
      "large language models generative": 17036,
      "llms demonstrated impressive performance": 18518,
      "demonstrated impressive performance various": 7530,
      "solving math word problems": 30512,
      "approach large language models": 2307,
      "large language models reasoning": 17193,
      "reasoning capabilities large language": 27385,
      "data recent advancements llms": 6828,
      "explore large language models": 10589,
      "recent success large language": 27561,
      "models llms recently showcased": 21613,
      "llms recently showcased remarkable": 18898,
      "language processing tasks diverse": 16801,
      "processing tasks diverse domains": 26128,
      "games large language models": 12082,
      "language models llms sparked": 16572,
      "language modeling question answering": 16225,
      "large language model inference": 16968,
      "language models llms exploded": 16469,
      "models llms exploded popularity": 21528,
      "pretrained language models contain": 25668,
      "tasks finetuning language models": 32338,
      "llms achieved impressive results": 18416,
      "advanced large language models": 1427,
      "language models llms pretrained": 16541,
      "benchmarking large language models": 3425,
      "task natural language processing": 32166,
      "llms demonstrated remarkable capabilities": 18522,
      "remarkable capabilities natural language": 28034,
      "capabilities natural language understanding": 4051,
      "achieve similar better performance": 986,
      "continual learning large language": 5977,
      "language models llms demonstrate": 16443,
      "models llms demonstrate exceptional": 21495,
      "acceleration large language models": 806,
      "sparse finetuning large language": 30614,
      "llms finetuning pretrained llms": 18613,
      "large language models chinese": 16999,
      "potential large language models": 25269,
      "large language models zeroshot": 17221,
      "language models llms gpt3": 16486,
      "capability large language models": 4095,
      "large language models gpt35": 17040,
      "large language models learn": 17051,
      "exceptional performance various tasks": 10173,
      "proprietary large language model": 26641,
      "available large language models": 2987,
      "large language models generation": 17035,
      "named entity recognition using": 22489,
      "language model llm using": 16171,
      "language models llms perform": 16535,
      "choice question answering mcqa": 4554,
      "llms smaller language models": 18957,
      "models llms chatgpt llama": 21489,
      "stateoftheart large language models": 30941,
      "opensource large language models": 23512,
      "models llms like llama": 21579,
      "address limitations present new": 1345,
      "conduct experiments diverse set": 5604,
      "using incontext learning icl": 34798,
      "recent advancements natural language": 27497,
      "natural language processing large": 22543,
      "language processing large language": 16785,
      "processing large language models": 26109,
      "work provides insights potential": 35767,
      "understanding large language models": 34241,
      "language models llms remarkable": 16552,
      "remains lack comprehensive investigation": 28000,
      "multilingual pretrained language models": 22325,
      "current landscape large language": 6498,
      "release code pretrained checkpoints": 27904,
      "challenging task natural language": 4401,
      "retrieval augmented large language": 28737,
      "commercial large language models": 4986,
      "evaluate effectiveness proposed methods": 9833,
      "language models language models": 16401,
      "models language models lms": 21418,
      "source domain target domains": 30559,
      "results natural language processing": 28651,
      "natural language processing computer": 22538,
      "language processing computer vision": 16779,
      "finetuned large language models": 11327,
      "large language models rapid": 17191,
      "language models rapid advancement": 16680,
      "rapid advancement large language": 27244,
      "advancement large language models": 1456,
      "various language models including": 35107,
      "language models llms including": 16500,
      "potential pretrained large language": 25291,
      "method large language models": 19939,
      "models llms shown great": 21630,
      "llms shown great potential": 18940,
      "great potential natural language": 13256,
      "potential natural language processing": 25283,
      "conduct comprehensive experiments demonstrate": 5595,
      "comprehensive experiments demonstrate effectiveness": 5378,
      "experiments demonstrate effectiveness method": 10433,
      "work provides valuable insights": 35769,
      "valuable insights future research": 35014,
      "systems using large language": 31926,
      "opensource models like llama": 23532,
      "like llama 7b 13b": 17884,
      "natural language processing task": 22555,
      "exhibited remarkable performance various": 10238,
      "llms including chatgpt gpt4": 18690,
      "quantization large language models": 27006,
      "pretrained language models study": 25694,
      "pretrained language models trained": 25695,
      "retrieval augmented generation rag": 28735,
      "direct preference optimization dpo": 8215,
      "cost training models scratch": 6257,
      "model 13 billion parameters": 20334,
      "work propose novel framework": 35758,
      "encoded large language models": 9342,
      "samples large language models": 29082,
      "open large language models": 23403,
      "applicability large language model": 2122,
      "large language models conduct": 17002,
      "including finetuning incontext learning": 14479,
      "pretrained language models llms": 25678,
      "models llms demonstrated superior": 21510,
      "significant progress natural language": 30012,
      "progress natural language processing": 26220,
      "using natural language explanations": 34851,
      "demonstrated large language models": 7534,
      "tasks incontext learning icl": 32371,
      "natural language explanations nles": 22517,
      "language models llms given": 16483,
      "proliferation large language models": 26260,
      "popular large language models": 25121,
      "popularity large language models": 25144,
      "recent advances transformerbased large": 27507,
      "advances transformerbased large language": 1490,
      "transformerbased large language models": 33755,
      "great strides natural language": 13261,
      "domains large language models": 8625,
      "models llms exhibit remarkable": 21526,
      "proprietary models gpt35 gpt4": 26649,
      "language models llms extensive": 16472,
      "deep neural network model": 7341,
      "source code publicly available": 30553,
      "number parameters language models": 23157,
      "teaching small language models": 32596,
      "small language models reason": 30349,
      "models llms demonstrated exceptional": 21498,
      "highperformance computing large language": 13678,
      "computing large language models": 5516,
      "models llms including llama": 21565,
      "responses response challenge propose": 28512,
      "generated qa questionanswer instances": 12381,
      "finetuning large language model": 11429,
      "natural language processing techniques": 22558,
      "language models paper present": 16644,
      "reasoning abilities large language": 27370,
      "language models conduct extensive": 16286,
      "models conduct extensive experiments": 21090,
      "conduct extensive experiments popular": 5608,
      "indicate significant performance gap": 14697,
      "language models llms llms": 16518,
      "language models llms various": 16590,
      "models llms various tasks": 21663,
      "survey large language models": 31776,
      "models llms gpt4 llama": 21548,
      "introduce novel inference method": 15529,
      "language models llms llama": 16517,
      "proximal policy optimization ppo": 26797,
      "language models llms increasingly": 16505,
      "large language models represent": 17198,
      "large language model meta": 16980,
      "language model meta ai": 16176,
      "advancement field natural language": 1449,
      "data source code publicly": 6873,
      "extend context window models": 10651,
      "policy gradient reinforcement learning": 25082,
      "large language models exploring": 17020,
      "code data model publicly": 4732,
      "data model publicly available": 6774,
      "local large language models": 19131,
      "processing nlp tasks inspired": 26122,
      "available apache 20 license": 2967,
      "large language models better": 16996,
      "learners large language models": 17528,
      "llms natural language understanding": 18812,
      "models llms highlights potential": 21556,
      "scales large language models": 29155,
      "models project page available": 21861,
      "representations large language models": 28164,
      "demonstrate proposed approach significantly": 7483,
      "models llms gpt4 llama2": 21549,
      "tasks natural language processing": 32429,
      "performance nonenglish languages paper": 24696,
      "question conduct extensive empirical": 27065,
      "development opensource large language": 7969,
      "advancing opensource language models": 1496,
      "sft direct preference optimization": 29762,
      "provide model finetuned follow": 26716,
      "model finetuned follow instructions": 20526,
      "models released apache 20": 21920,
      "released apache 20 license": 27921,
      "llms gpt35 gpt4 llama2": 18652,
      "multimodal large language models": 22355,
      "language models llms multimodal": 16522,
      "models llms multimodal large": 21586,
      "llms multimodal large language": 18807,
      "large language models mllms": 17173,
      "language models mllms shown": 16626,
      "large language models mental": 17170,
      "language models mental health": 16619,
      "models like bert xlnet": 21450,
      "language models llms potential": 16537,
      "performance downstream tasks paper": 24576,
      "downstream tasks paper explore": 8698,
      "language models era large": 16323,
      "models era large language": 21214,
      "language models llms hold": 16495,
      "instruction tuning large language": 15198,
      "tuning large language models": 33989,
      "models llms demonstrated impressive": 21501,
      "llms demonstrated impressive capabilities": 18517,
      "using reinforcement learning rl": 34902,
      "language models improve performance": 16383,
      "parameter efficient finetuning peft": 24181,
      "processing nlp tasks paper": 26123,
      "models llms excel tasks": 21523,
      "despite general capabilities large": 7779,
      "large language models study": 17205,
      "language models language model": 16400,
      "iterations approach yields model": 15685,
      "approach yields model outperforms": 2366,
      "large language models scientific": 17200,
      "language models llms task": 16579,
      "fusion large language models": 12024,
      "large language models training": 17212,
      "language models training large": 16736,
      "models training large language": 22080,
      "code model weights data": 4784,
      "model weights data public": 20868,
      "empirical results demonstrate proposed": 9235,
      "future research practical applications": 12049,
      "explainability large language models": 10528,
      "language models llms critical": 16441,
      "natural language processing llms": 22545,
      "pretraining large language models": 25812,
      "reasoning multimodal large language": 27427,
      "multimodal large language modelsmllms": 22359,
      "extreme compression large language": 10793,
      "compression large language models": 5419,
      "size poses significant challenges": 30273,
      "multilingual capabilities large language": 22299,
      "extending large language models": 10665,
      "recent advances natural language": 27504,
      "advances natural language processing": 1485,
      "finance large language models": 11212,
      "versatile multimodal large language": 35226,
      "multimodal large language model": 22353,
      "large language model mllm": 16982,
      "demonstrates significant performance improvements": 7569,
      "language models nlp tasks": 16635,
      "aligning large language models": 1747,
      "large language models news": 17178,
      "data training evaluation code": 6898,
      "llms llama2 gpt35 palm2": 18776,
      "llms 7b 70b parameters": 18404,
      "large language models various": 17217,
      "language models capable performing": 16272,
      "based large language models": 3188,
      "language models llms particularly": 16534,
      "substantial computational memory requirements": 31463,
      "guardrails large language models": 13337,
      "parameterefficient finetuning peft methods": 24210,
      "commonsense reasoning reading comprehension": 5043,
      "language models llms proven": 16544,
      "task large language models": 32150,
      "sparsity large language models": 30632,
      "natural approach reduce cost": 22505,
      "llms like gpt llama": 18760,
      "strong performance wide range": 31187,
      "compare performance popular llms": 5113,
      "open challenges future research": 23388,
      "models diverse set tasks": 21166,
      "large language model recent": 16985,
      "large language models achieved": 16990,
      "finetuned llama model significantly": 11332,
      "llama model significantly outperforms": 18131,
      "language models llms great": 16492,
      "models llms increasingly utilized": 21569,
      "applied large language models": 2193,
      "learning natural language processing": 17621,
      "natural language processing based": 22537,
      "language model training data": 16209,
      "introduces novel evaluation framework": 15547,
      "evaluation framework large language": 9950,
      "framework large language models": 11876,
      "open source large language": 23424,
      "large language model llama2": 16971,
      "transformerbased large language model": 33753,
      "prompts large language models": 26428,
      "challenges large language models": 4356,
      "gap introduce new benchmark": 12093,
      "training language models lms": 33541,
      "language models lms strong": 16608,
      "language models llms knowledge": 16508,
      "llms knowledge graphs kgs": 18737,
      "code data publicly released": 4739,
      "language models llms new": 16527,
      "language models llms make": 16519,
      "incontext learning icl capabilities": 14557,
      "graph neural network gnn": 13227,
      "models llms demonstrated strong": 21509,
      "large language models gpt4": 17041,
      "languages large language models": 16886,
      "usage large language models": 34507,
      "language models lms proven": 16606,
      "llms pretrained large language": 18866,
      "efficiency large language models": 9007,
      "models llms demonstrate remarkable": 21496,
      "study large language models": 31354,
      "pretrained language models improving": 25674,
      "yields significant performance gains": 35929,
      "tokens large language models": 33239,
      "remarkable capabilities various tasks": 28039,
      "significant advancement field natural": 29952,
      "evaluate large language models": 9843,
      "novel approach designed reduce": 23058,
      "models available hugging face": 21003,
      "language models explore approach": 16336,
      "language models plms bert": 16654,
      "language models llms ability": 16426,
      "time large language models": 33132,
      "text generation large language": 32873,
      "language models llms known": 16509,
      "propose novel evaluation framework": 26548,
      "language models achieved remarkable": 16239,
      "models achieved remarkable success": 20951,
      "general language understanding tasks": 12176,
      "language adaptation large language": 16037,
      "adaptation large language models": 1182,
      "large language models including": 17045,
      "paper try answer question": 24145,
      "tasks maintaining comparable performance": 32414,
      "language models recent advances": 16686,
      "hallucination code data available": 13375,
      "instruction data evaluation benchmark": 15145,
      "language models minimal human": 16623,
      "serving large language models": 29666,
      "language models recent works": 16688,
      "future large language models": 12037,
      "large language models recently": 17197,
      "large language models potentially": 17185,
      "superior zeroshot performance new": 31660,
      "language models llms stand": 16574,
      "compared 350m parameter opt": 5120,
      "large language models specifically": 17203,
      "human feedback rlhf framework": 13822,
      "language models llms specifically": 16573,
      "language models llms use": 16585,
      "pretrained language models using": 25696,
      "model code data available": 20424,
      "models like gpt35 llama2": 21456,
      "rapid advancement generative artificial": 27241,
      "advancement generative artificial intelligence": 1453,
      "generative artificial intelligence genai": 12657,
      "generative large language model": 12668,
      "language model llm inference": 16168,
      "instructionfinetuned large language models": 15220,
      "emergence numerous large language": 9177,
      "numerous large language models": 23186,
      "language processing nlp applications": 16791,
      "properties large language models": 26476,
      "improves downstream task performance": 14374,
      "large language models explore": 17019,
      "contemporary large language models": 5844,
      "language models billions parameters": 16266,
      "large language models fewshot": 17021,
      "widespread adoption large language": 35593,
      "adoption large language models": 1410,
      "language models llms facilitated": 16473,
      "language models llms understanding": 16584,
      "various llms including gpt4": 35115,
      "llms including gpt4 llama2": 18701,
      "generative artificial intelligence gai": 12656,
      "language models lms various natural": 16612,
      "models lms various natural language": 21685,
      "lms various natural language processing": 19125,
      "various natural language processing tasks": 35128,
      "language models large language models": 16406,
      "models large pretrained language models": 21428,
      "generative pretrained language model gpt2": 12688,
      "bidirectional encoder representations transformers bert": 3695,
      "gpt2 radford et al 2019": 12945,
      "language models lms demonstrated impressive": 16600,
      "pretrained language models recent years": 25691,
      "size pretrained language models plms": 30277,
      "stateoftheart results various natural language": 30989,
      "various natural language processing nlp": 35127,
      "natural language processing nlp tasks": 22554,
      "natural language understanding generation tasks": 22571,
      "natural language processing nlp domain": 22550,
      "language models pretrained language models": 16663,
      "wide range natural language processing": 35559,
      "range natural language processing nlp": 27202,
      "recent years pretrained language models": 27581,
      "pretrained language models bert gpt2": 25666,
      "bert roberta gpt2 dozens datasets": 3531,
      "language models large pretrained language": 16409,
      "make code models publicly available": 19457,
      "remarkable success large language models": 28059,
      "machine learning models large language": 19351,
      "models large language models llm": 21423,
      "radford et al 2019 gpt3": 27153,
      "et al 2019 gpt3 brown": 9794,
      "al 2019 gpt3 brown et": 1685,
      "2019 gpt3 brown et al": 226,
      "gpt3 brown et al 2020": 12986,
      "natural language processing nlp algorithms": 22548,
      "automated natural language generation metrics": 2873,
      "berts masked language modeling mlm": 3549,
      "question answering natural language inference": 27050,
      "challenge natural language processing nlp": 4322,
      "natural language processing nlp systems": 22553,
      "natural language understanding nlu natural": 22573,
      "language understanding nlu natural language": 16846,
      "understanding nlu natural language generation": 34257,
      "nlu natural language generation nlg": 22974,
      "generative pretrained language models plms": 12690,
      "benefit using large language models": 3487,
      "using large language models llms": 34812,
      "natural language understanding nlu tasks": 22574,
      "recent large language models llms": 27527,
      "large language models llms demonstrated": 17069,
      "language models llms demonstrated remarkable": 16450,
      "models large language models llms": 21424,
      "large language models large language": 17050,
      "large language models llms shown": 17145,
      "improve model performance generalization unseen": 14278,
      "model performance generalization unseen tasks": 20702,
      "models long short term memory": 21690,
      "language model large language models": 16161,
      "model large language models llms": 20605,
      "large language models llms led": 17109,
      "landscape large language models llms": 16026,
      "largescale pretrained language models bert": 17377,
      "knowledge large language models llms": 15874,
      "large language models llms trained": 17156,
      "enables pretrained language models perform": 9309,
      "large language models llms surprisingly": 17153,
      "safety large language models llms": 29051,
      "performance various natural language processing": 24804,
      "summarization large language models llms": 31617,
      "large language models llms used": 17162,
      "recently large language models llms": 27609,
      "pretrained language models like bert": 25677,
      "large neural language models trained": 17249,
      "use large language models llms": 34545,
      "large language models llms chatgpt": 17064,
      "data large language models llms": 6752,
      "large language models llms downstream": 17073,
      "large language models llms revolutionized": 17142,
      "language models llms revolutionized natural": 16557,
      "models llms revolutionized natural language": 21625,
      "llms revolutionized natural language processing": 18926,
      "large language models llms gpt4": 17096,
      "success large language models llms": 31517,
      "large language models llms like": 17110,
      "language models llms like gpt4": 16515,
      "feedback large language models llms": 11067,
      "language models llms like chatgpt": 16513,
      "make model data code publicly": 19475,
      "instruction following large language model": 15168,
      "research field natural language processing": 28315,
      "adapting large language models llms": 1211,
      "deploying large language models llms": 7642,
      "large language models llms challenging": 17063,
      "ability large language models llms": 692,
      "descriptions large language models llms": 7691,
      "large language models llms multiple": 17116,
      "large language models follow instructions": 17027,
      "large language models llms produce": 17133,
      "harnessing capabilities large language models": 13462,
      "generative large language models llms": 12670,
      "language models llms revolutionized field": 16556,
      "comes significant computational costs paper": 4974,
      "proprietary large language models llms": 26644,
      "large pretrained language models demonstrated": 17264,
      "large language models llms directly": 17072,
      "text generated large language models": 32864,
      "leveraging large language models llms": 17786,
      "large language models llms data": 17067,
      "comprehensive evaluation large language models": 5372,
      "make data code publicly available": 19464,
      "tasks large language models llms": 32393,
      "large language models recent advancements": 17195,
      "language models recent advancements large": 16685,
      "models recent advancements large language": 21904,
      "recent advancements large language models": 27494,
      "advancements large language models llms": 1468,
      "large language models llms significantly": 17147,
      "learning large language models llms": 17595,
      "large language models llms acquire": 17060,
      "excel various natural language processing": 10158,
      "large language models llms study": 17152,
      "language models large language modelsllms": 16407,
      "entities pretrained language models lms": 9638,
      "large language models llms achieved": 17059,
      "language models llms achieved great": 16429,
      "models llms achieved great success": 21477,
      "large language model llm based": 16973,
      "recent years large language models": 27578,
      "emergence large language models llms": 9172,
      "pretrained language model plm t5": 25661,
      "large language model based llama": 16963,
      "large language models llms capture": 17062,
      "models llms demonstrated remarkable potential": 21507,
      "experimental results demonstrate superior performance": 10398,
      "nlp tasks large language models": 22959,
      "large language models llms typically": 17159,
      "developments large language models llms": 7983,
      "large language models llms enabled": 17077,
      "pretrained large language models llms": 25701,
      "large language models llms recently": 17136,
      "language models llms recently achieved": 16547,
      "generation large language models llms": 12537,
      "large language models llms widely": 17166,
      "language models llms shown remarkable": 16566,
      "effectiveness various generaldomain natural language": 8974,
      "various generaldomain natural language processing": 35099,
      "generaldomain natural language processing nlp": 12194,
      "development large language models llms": 7960,
      "supervised finetuning sft reinforcement learning": 31681,
      "finetuning sft reinforcement learning human": 11523,
      "sft reinforcement learning human feedback": 29767,
      "reinforcement learning human feedback rlhf": 27844,
      "results various natural language tasks": 28708,
      "reasoning large language models llms": 27418,
      "large language models llms exhibit": 17081,
      "great success large language models": 13265,
      "recent advances large language models": 27502,
      "advances large language models llm": 1481,
      "instructionfollowing large language models llms": 15233,
      "models pretrained large language models": 21845,
      "supervised finetuning reinforcement learning human": 31678,
      "finetuning reinforcement learning human feedback": 11506,
      "diverse natural language processing tasks": 8444,
      "advanced natural language processing nlp": 1436,
      "natural language processing nlp research": 22552,
      "instructions large language models llms": 15259,
      "large language models llms present": 17129,
      "revolutionized field natural language processing": 28847,
      "field natural language processing enabling": 11146,
      "significant advancements natural language processing": 29957,
      "models range natural language processing": 21888,
      "range natural language processing tasks": 27203,
      "recent progress large language models": 27544,
      "progress large language models llms": 26215,
      "technical report large language models": 32610,
      "report large language models llms": 28120,
      "language models llms achieved remarkable": 16430,
      "models llms achieved remarkable success": 21480,
      "large language models despite impressive": 17009,
      "despite impressive capabilities large language": 7791,
      "impressive capabilities large language models": 14234,
      "capabilities large language models llms": 4035,
      "language models llms gpt4 palm": 16491,
      "models llms gpt4 palm llama": 21551,
      "finetuning peftlora based approach used": 11481,
      "peftlora based approach used study": 24443,
      "based approach used study model": 3136,
      "approach used study model finetuned": 2354,
      "used study model finetuned following": 34627,
      "study model finetuned following tasks": 31365,
      "model finetuned following tasks analysing": 20529,
      "finetuned following tasks analysing text": 11315,
      "sentiments obtained results finetuned llama": 29581,
      "obtained results finetuned llama model": 23260,
      "results finetuned llama model perform": 28616,
      "extracted sentiments named entities considered": 10755,
      "sentiments named entities considered predictive": 29577,
      "named entities considered predictive features": 22484,
      "entities considered predictive features supervised": 9633,
      "considered predictive features supervised machine": 5721,
      "predictive features supervised machine learning": 25452,
      "features supervised machine learning models": 11042,
      "existing large language models llms": 10286,
      "large language models llms generate": 17089,
      "large language models llms gpt": 17093,
      "large language models llms need": 17118,
      "large language models llms excel": 17080,
      "language models llms excel various": 16466,
      "inference large language models llms": 14784,
      "revolutionized natural language processing nlp": 28851,
      "models llms demonstrated remarkable performance": 21506,
      "enhance capabilities large language models": 9508,
      "machine translation large language models": 19358,
      "neural machine translation nmt systems": 22734,
      "large language models llms emerged": 17075,
      "large language models llms represent": 17140,
      "best performing models achieved accuracy": 3573,
      "autoregressive large language models llms": 2950,
      "reasoning ability llms large language": 27376,
      "ability llms large language models": 702,
      "llms large language models llms": 18746,
      "large language models llms gained": 17087,
      "language models llms shown promise": 16565,
      "capabilities natural language processing nlp": 4050,
      "reasoning capabilities large language models": 27386,
      "explore large language models llms": 10590,
      "recent success large language models": 27562,
      "language models llms recently showcased": 16548,
      "models llms recently showcased remarkable": 21614,
      "natural language processing tasks diverse": 22557,
      "language processing tasks diverse domains": 16802,
      "games large language models llms": 12083,
      "advances large language models llms": 1482,
      "large language models llms sparked": 17149,
      "large language models llms exploded": 17082,
      "language models llms exploded popularity": 16470,
      "large language models llms pretrained": 17131,
      "using large language models large": 34811,
      "models llms demonstrated remarkable capabilities": 21505,
      "remarkable capabilities natural language understanding": 28036,
      "continual learning large language models": 5978,
      "large language models llms demonstrate": 17068,
      "language models llms demonstrate exceptional": 16444,
      "sparse finetuning large language models": 30615,
      "finetuning large language models llms": 11432,
      "potential large language models llms": 25270,
      "large language models llms gpt3": 17094,
      "proprietary large language model llm": 26642,
      "instructiontuned large language models llms": 15291,
      "instructiontuned large language model llm": 15289,
      "large language model llm using": 16978,
      "large language models llms perform": 17126,
      "language models llms chatgpt llama": 16438,
      "stateoftheart large language models large": 30942,
      "opensource large language models llms": 23513,
      "language models llms like llama": 16516,
      "recent advancements natural language processing": 27498,
      "advancements natural language processing large": 1473,
      "natural language processing large language": 22544,
      "language processing large language models": 16786,
      "processing large language models llms": 26110,
      "understanding large language models large": 34242,
      "large language models llms remarkable": 17139,
      "current landscape large language models": 6499,
      "challenging task natural language processing": 4402,
      "commercial large language models llms": 4987,
      "language models language models lms": 16402,
      "natural language processing computer vision": 22539,
      "revolutionized natural language processing tasks": 28852,
      "evaluating large language models llms": 9903,
      "large language models rapid advancement": 17192,
      "rapid advancement large language models": 27245,
      "advancement large language models llms": 1457,
      "large language models llms including": 17104,
      "potential pretrained large language models": 25292,
      "method large language models llms": 19940,
      "language models llms shown great": 16562,
      "models llms shown great potential": 21631,
      "great potential natural language processing": 13257,
      "potential natural language processing nlp": 25284,
      "systems using large language models": 31927,
      "quantization large language models llms": 27007,
      "large pretrained language models llms": 17265,
      "language models llms demonstrated superior": 16453,
      "research large language models llms": 28333,
      "significant progress natural language processing": 30013,
      "progress natural language processing nlp": 26221,
      "demonstrated large language models llms": 7535,
      "large language models llms given": 17092,
      "proliferation large language models llms": 26261,
      "recent advances transformerbased large language": 27508,
      "domains large language models llms": 8626,
      "language models llms exhibit remarkable": 16468,
      "training large language models llms": 33546,
      "large language models llms extensive": 17083,
      "teaching small language models reason": 32597,
      "language models llms demonstrated exceptional": 16447,
      "highperformance computing large language models": 13679,
      "computing large language models llms": 5517,
      "language models llms including llama": 16503,
      "large language models paper present": 17181,
      "reasoning abilities large language models": 27371,
      "large language models conduct extensive": 17003,
      "language models conduct extensive experiments": 16287,
      "models conduct extensive experiments popular": 21091,
      "multilingual large language models llms": 22316,
      "large language models llms llms": 17112,
      "large language models llms various": 17165,
      "language models llms various tasks": 16591,
      "language models llms gpt4 llama": 16489,
      "large language models llms llama": 17111,
      "family large language models llms": 10978,
      "generated large language models llms": 12370,
      "large language models llms increasingly": 17106,
      "large language model meta ai": 16981,
      "advancement field natural language processing": 1450,
      "data source code publicly available": 6874,
      "code data model publicly available": 4733,
      "especially large language models llms": 9742,
      "language processing nlp tasks inspired": 16797,
      "language models llms gpt4 llama2": 16490,
      "development opensource large language models": 7970,
      "sft direct preference optimization dpo": 29763,
      "provide model finetuned follow instructions": 26717,
      "models released apache 20 license": 21921,
      "capability large language models llms": 4096,
      "multimodal large language models large": 22356,
      "large language models llms multimodal": 17115,
      "language models llms multimodal large": 16523,
      "models llms multimodal large language": 21587,
      "llms multimodal large language models": 18808,
      "multimodal large language models mllms": 22358,
      "large language models mllms shown": 17175,
      "large language models mental health": 17171,
      "large language models llms potential": 17128,
      "language models era large language": 16324,
      "models era large language models": 21215,
      "analysis large language models llms": 1937,
      "large language models llms hold": 17100,
      "instruction tuning large language models": 15199,
      "tuning large language models llms": 33990,
      "language models llms demonstrated impressive": 16449,
      "models llms demonstrated impressive capabilities": 21502,
      "performance large language models llms": 24648,
      "language processing nlp tasks paper": 16798,
      "language models llms excel tasks": 16465,
      "applications large language models llms": 2162,
      "iterations approach yields model outperforms": 15686,
      "open large language models llms": 23404,
      "large language models llms task": 17155,
      "language models training large language": 16737,
      "models training large language models": 22081,
      "code model weights data public": 4785,
      "explainability large language models llms": 10529,
      "large language models llms critical": 17066,
      "advanced large language models llms": 1428,
      "extreme compression large language models": 10794,
      "multilingual capabilities large language models": 22300,
      "extending large language models llms": 10666,
      "recent advances natural language processing": 27505,
      "finance large language models llms": 11213,
      "versatile multimodal large language model": 35227,
      "multimodal large language model mllm": 22354,
      "pretrained language models nlp tasks": 25681,
      "large pretrained language models plms": 17266,
      "based large language models llms": 3189,
      "large language models llms particularly": 17125,
      "large language models llms proven": 17134,
      "years large language models achieved": 35894,
      "finetuned llama model significantly outperforms": 11333,
      "large language models llms great": 17097,
      "datasets large language models llms": 7141,
      "language models llms increasingly utilized": 16507,
      "applied large language models llms": 2194,
      "evaluation framework large language models": 9951,
      "transformerbased large language model llm": 33754,
      "prompts large language models llms": 26429,
      "large language models llms knowledge": 17107,
      "large language models llms new": 17119,
      "current large language models llms": 6505,
      "large language models llms make": 17113,
      "language models llms demonstrated strong": 16452,
      "languages large language models llms": 16887,
      "usage large language models llms": 34508,
      "llms pretrained large language models": 18867,
      "language models llms demonstrate remarkable": 16445,
      "evaluation large language models llms": 9968,
      "study large language models llms": 31355,
      "significant advancement field natural language": 29953,
      "field natural language processing nlp": 11147,
      "models llms demonstrated impressive performance": 21503,
      "pretrained language models plms bert": 25686,
      "text generation large language models": 32874,
      "large language models llms known": 17108,
      "large language models achieved remarkable": 16991,
      "language models achieved remarkable success": 16240,
      "language adaptation large language models": 16038,
      "large language models recent works": 17196,
      "using large language models recently": 34813,
      "effectiveness large language models llms": 8953,
      "large language models llms stand": 17151,
      "learning human feedback rlhf framework": 17580,
      "large language models llms specifically": 17150,
      "large language models llms use": 17161,
      "scaling large language models llms": 29170,
      "task large language models llms": 32151,
      "abilities large language models llms": 631,
      "rapid advancement generative artificial intelligence": 27242,
      "large language model llm inference": 16975,
      "emergence numerous large language models": 9178,
      "natural language processing nlp applications": 22549,
      "contemporary large language models llms": 5845,
      "performance various natural language tasks": 24805,
      "stateoftheart large language models llms": 30943,
      "widespread adoption large language models": 35594,
      "adoption large language models llms": 1411,
      "large language models llms facilitated": 17084,
      "large language models llms understanding": 17160,
      "bypasses": 3950,
      "byte": 3954,
      "retrained": 28725,
      "distractor": 8381,
      "grover": 13308,
      "infinitely": 14836,
      "contextualizing": 5965,
      "selfsimilarity": 29435,
      "justification": 15735,
      "electronic": 9097,
      "auc": 2771,
      "alternating": 1847,
      "priors": 25949,
      "speaker": 30648,
      "persuasion": 24902,
      "persuade": 24901,
      "traverse": 33870,
      "moving": 22239,
      "fosters": 11789,
      "fits": 11573,
      "paragraphs": 24164,
      "shortened": 29824,
      "morphological": 22219,
      "tiling": 33106,
      "concluded": 5553,
      "welldocumented": 35522,
      "associating": 2654,
      "priori": 25944,
      "dual": 8750,
      "injects": 14984,
      "manifested": 19537,
      "adhoc": 1387,
      "jensenshannon": 15703,
      "organic": 23688,
      "rough": 28984,
      "plots": 25054,
      "completed": 5253,
      "hierarchy": 13547,
      "deciding": 7229,
      "inform": 14849,
      "095": 22,
      "accompanying": 839,
      "83b": 572,
      "conditionally": 5574,
      "covid19": 6332,
      "pandemic": 24000,
      "calling": 3983,
      "june": 15732,
      "74": 512,
      "forums": 11780,
      "venues": 35200,
      "post": 25220,
      "members": 19789,
      "posted": 25222,
      "retrain": 28724,
      "cord19": 6148,
      "shot": 29830,
      "spacing": 30587,
      "corner": 6161,
      "bioinformatics": 3741,
      "protein": 26659,
      "electra": 9096,
      "tpu": 33321,
      "1024": 56,
      "localization": 19138,
      "bypassing": 3951,
      "318": 317,
      "connected": 5682,
      "dp": 8706,
      "differently": 8165,
      "priming": 25927,
      "implication": 14171,
      "ending": 9426,
      "interdependency": 15402,
      "lite": 18036,
      "diverge": 8405,
      "spawning": 30646,
      "distortion": 8379,
      "confounding": 5670,
      "ranker": 27227,
      "xai": 35867,
      "visualize": 35362,
      "wechat": 35490,
      "hot": 13762,
      "tries": 33897,
      "elicited": 9104,
      "enjoys": 9578,
      "verifiable": 35208,
      "treating": 33873,
      "regimes": 27817,
      "insertion": 15060,
      "peak": 24429,
      "50k": 427,
      "rankers": 27228,
      "similaritybased": 30132,
      "thresholds": 33097,
      "09": 20,
      "concatenates": 5521,
      "averages": 3030,
      "distractors": 8382,
      "missed": 20222,
      "presumably": 25620,
      "chats": 4508,
      "autocomplete": 2840,
      "shares": 29789,
      "unintended": 34351,
      "wisdom": 35621,
      "tagger": 32010,
      "transformersbased": 33799,
      "chemical": 4526,
      "drug": 8747,
      "cleaned": 4644,
      "affective": 1543,
      "category": 4231,
      "insignificant": 15085,
      "captions": 4141,
      "confront": 5671,
      "bar": 3094,
      "requested": 28207,
      "conceptually": 5537,
      "pseudo": 26818,
      "letter": 17727,
      "regarded": 27807,
      "satisfies": 29109,
      "constraining": 5783,
      "treatment": 33874,
      "backtranslation": 3069,
      "clip": 4662,
      "visionandlanguage": 35311,
      "reconstructed": 27659,
      "fault": 11006,
      "clickthrough": 4650,
      "enjoyable": 9576,
      "lewis": 17796,
      "179": 180,
      "agree": 1589,
      "encodings": 9384,
      "blanks": 3759,
      "reasoners": 27366,
      "qnli": 26925,
      "mnli": 20317,
      "isnt": 15648,
      "compete": 5210,
      "pc": 24427,
      "lowers": 19294,
      "bigru": 3704,
      "comment": 4980,
      "evil": 10071,
      "vl": 35373,
      "wellsuited": 35530,
      "hearing": 13489,
      "paucity": 24415,
      "117": 81,
      "gb": 12134,
      "diminish": 8201,
      "hans": 13416,
      "diagnostics": 8002,
      "scene": 29223,
      "cross": 6404,
      "endowing": 9428,
      "ample": 1892,
      "tackled": 32003,
      "qualities": 26937,
      "intersectional": 15471,
      "subgroups": 31422,
      "tokenbytoken": 33206,
      "drift": 8733,
      "bidirectionally": 3697,
      "logs": 19164,
      "probable": 25974,
      "optimality": 23618,
      "converges": 6087,
      "inputting": 15053,
      "taming": 32042,
      "beast": 3301,
      "extendable": 10658,
      "removal": 28073,
      "unsatisfactory": 34430,
      "entropybased": 9662,
      "unifiedqa": 34341,
      "realizes": 27322,
      "barely": 3101,
      "discard": 8254,
      "unit": 34366,
      "theories": 33055,
      "overcomes": 23926,
      "standardization": 30884,
      "accounting": 852,
      "disclosure": 8261,
      "engaged": 9453,
      "swapping": 31790,
      "deteriorates": 7891,
      "transferable": 33686,
      "weaklysupervised": 35462,
      "125": 95,
      "impose": 14220,
      "guaranteeing": 13331,
      "appearance": 2114,
      "speaking": 30650,
      "truefalse": 33926,
      "blind": 3772,
      "boolq": 3813,
      "t5s": 31977,
      "kd": 15744,
      "prize": 25960,
      "music": 22469,
      "tv": 34034,
      "orion": 23731,
      "acknowledge": 1111,
      "favor": 11009,
      "propagate": 26464,
      "competent": 5213,
      "immediate": 14111,
      "dramatic": 8716,
      "selfalignment": 29405,
      "abundant": 780,
      "fused": 12018,
      "domainadapted": 8605,
      "prefixed": 25482,
      "forced": 11717,
      "winogrande": 35618,
      "facial": 10833,
      "disclose": 8260,
      "metalearning": 19861,
      "evades": 9814,
      "privacypreserving": 25955,
      "cotraining": 6286,
      "crossattention": 6406,
      "gpt2xl": 12978,
      "datafree": 6930,
      "interconnected": 15400,
      "interplay": 15451,
      "wav2vec20": 35426,
      "doesnt": 8544,
      "invariant": 15564,
      "isolate": 15649,
      "uncommon": 34129,
      "430": 389,
      "valence": 34987,
      "correspondence": 6223,
      "rho": 28872,
      "recovered": 27668,
      "weat": 35468,
      "intralayer": 15481,
      "spearmans": 30653,
      "reader": 27298,
      "diagnosis": 8000,
      "governed": 12838,
      "generalizations": 12230,
      "protected": 26655,
      "humanevaluation": 13898,
      "linguistically": 18025,
      "drawback": 8724,
      "hurt": 13951,
      "markov": 19603,
      "retrievalaugmentation": 28759,
      "nonretrieval": 23006,
      "united": 34367,
      "570": 450,
      "090": 21,
      "lie": 17818,
      "flaws": 11614,
      "copied": 6141,
      "penalize": 24445,
      "cope": 6139,
      "influences": 14844,
      "sentential": 29564,
      "higherlevel": 13614,
      "manifold": 19538,
      "biological": 3742,
      "restricting": 28533,
      "selfsupervision": 29442,
      "paretofrontier": 24310,
      "recommender": 27655,
      "usergenerated": 34681,
      "exiting": 10332,
      "unambiguous": 34105,
      "routinely": 28993,
      "humanintheloop": 13901,
      "perturbation": 24908,
      "errorprone": 9719,
      "496": 411,
      "bartlarge": 3112,
      "366": 355,
      "nonparametric": 23004,
      "perplexitybased": 24874,
      "terrible": 32755,
      "silly": 30092,
      "locally": 19143,
      "counterintuitive": 6299,
      "consecutive": 5692,
      "selfreinforcement": 29432,
      "bm25": 3800,
      "meetings": 19783,
      "phone": 24934,
      "oral": 23664,
      "translators": 33859,
      "conquered": 5691,
      "composable": 5318,
      "differential": 8162,
      "edit": 8822,
      "distinguishing": 8377,
      "assists": 2641,
      "barrier": 3102,
      "lab": 15950,
      "deductively": 7311,
      "mediumsized": 19777,
      "tells": 32685,
      "alleviating": 1795,
      "imagetotext": 14096,
      "offered": 23295,
      "ungrounded": 34316,
      "constrain": 5777,
      "dividing": 8491,
      "reformulated": 27800,
      "datastore": 7193,
      "exploitation": 10556,
      "39x": 363,
      "spikes": 30807,
      "titan": 33176,
      "int4": 15313,
      "2080": 244,
      "ti": 33101,
      "bug": 3904,
      "bugs": 3906,
      "167": 162,
      "84": 573,
      "gradientguided": 13197,
      "delays": 7393,
      "forces": 11718,
      "battery": 3292,
      "normal": 23016,
      "analytically": 1984,
      "transferlearning": 33688,
      "reinforced": 27836,
      "computeefficient": 5500,
      "convolution": 6121,
      "parser": 24316,
      "115": 79,
      "inherits": 14956,
      "commons": 5033,
      "cd": 4261,
      "opt13b": 23604,
      "opt125m": 23602,
      "subjectverb": 31428,
      "assert": 2585,
      "diffusionbased": 8186,
      "hubert": 13777,
      "explorations": 10566,
      "machinetranslated": 19374,
      "196": 193,
      "492": 410,
      "empowered": 9267,
      "reasonings": 27466,
      "prefers": 25478,
      "subproblems": 31443,
      "marketing": 19599,
      "fasttext": 11005,
      "classified": 4622,
      "inputoutput": 15038,
      "turing": 34025,
      "undetectable": 34294,
      "fooling": 11711,
      "mechanics": 19747,
      "clarity": 4580,
      "delivery": 7403,
      "unanswered": 34108,
      "denote": 7602,
      "flant5s": 11604,
      "specifics": 30765,
      "ast": 2666,
      "crossmodal": 6423,
      "neglected": 22667,
      "confused": 5673,
      "workarounds": 35798,
      "80m": 559,
      "soda": 30446,
      "standing": 30889,
      "exceptionally": 10176,
      "ushered": 34718,
      "2500": 272,
      "buying": 3946,
      "idiosyncratic": 14032,
      "compound": 5331,
      "30b": 313,
      "relevancy": 27934,
      "flaw": 11612,
      "ubiquitously": 34089,
      "imposed": 14221,
      "minilm": 20179,
      "ms": 22248,
      "drama": 8715,
      "outlines": 23744,
      "scenes": 29225,
      "templatebased": 32690,
      "investors": 15619,
      "profit": 26190,
      "lexicons": 17806,
      "formula": 11768,
      "achievable": 940,
      "decreased": 7299,
      "informationtheoretic": 14930,
      "functionally": 11968,
      "16k": 167,
      "prevalently": 25857,
      "incompleteness": 14538,
      "assurance": 2665,
      "tedious": 32681,
      "pressures": 25619,
      "localizes": 19141,
      "communications": 5053,
      "productive": 26171,
      "arc": 2432,
      "amt": 1900,
      "manifest": 19535,
      "attracting": 2753,
      "universities": 34375,
      "wait": 35406,
      "byt5": 3953,
      "bytelevel": 3955,
      "recommend": 27647,
      "prioritize": 25946,
      "chai": 4288,
      "translates": 33815,
      "nextgeneration": 22891,
      "epistemic": 9670,
      "normalized": 23019,
      "universality": 34372,
      "webpages": 35484,
      "consumers": 5820,
      "swiftly": 31792,
      "finer": 11279,
      "incapable": 14441,
      "endeavor": 9423,
      "laborious": 15972,
      "astray": 2669,
      "52k": 438,
      "schemas": 29231,
      "fulfill": 11944,
      "evenly": 10044,
      "nonuniform": 23014,
      "trap": 33867,
      "elastic": 9095,
      "contradict": 6005,
      "fisher": 11568,
      "informationseeking": 14929,
      "channel": 4418,
      "communicative": 5054,
      "cooperative": 6134,
      "opensourcing": 23557,
      "ignores": 14039,
      "reorder": 28084,
      "nonexperts": 22999,
      "videotext": 35269,
      "clips": 4669,
      "ocr": 23279,
      "optimally": 23619,
      "alpha": 1841,
      "intractable": 15480,
      "simplest": 30169,
      "multimodel": 22371,
      "youtube": 35930,
      "personally": 24891,
      "minigpt4": 20176,
      "humorous": 13941,
      "poems": 25060,
      "remote": 28072,
      "dino": 8208,
      "profoundly": 26192,
      "benchmarked": 3420,
      "restful": 28528,
      "costfree": 6262,
      "convenience": 6069,
      "batches": 3291,
      "phrasing": 24939,
      "criterion": 6379,
      "pinpoints": 24958,
      "indirect": 14709,
      "leak": 17497,
      "unimodal": 34349,
      "understands": 34279,
      "boosted": 3818,
      "4times": 415,
      "crosslanguage": 6411,
      "files": 11163,
      "converts": 6115,
      "multiround": 22436,
      "visuallanguage": 35365,
      "fulfilling": 11945,
      "submodules": 31436,
      "underwent": 34289,
      "impressively": 14251,
      "render": 28078,
      "artists": 2546,
      "longhorizon": 19209,
      "assistive": 2640,
      "executable": 10188,
      "robotics": 28927,
      "polarities": 25076,
      "contextualised": 5958,
      "usages": 34510,
      "exemplar": 10203,
      "singlegpu": 30231,
      "distinguished": 8375,
      "tackles": 32004,
      "spreading": 30829,
      "continuity": 5994,
      "68": 485,
      "faulty": 11008,
      "mood": 22217,
      "224": 254,
      "237": 258,
      "max": 19698,
      "modelslms": 22147,
      "guessing": 13339,
      "134x": 115,
      "122": 93,
      "nearperfect": 22603,
      "multidigit": 22274,
      "breadth": 3857,
      "prioritizes": 25947,
      "stopping": 31074,
      "hallucinates": 13371,
      "cka": 4575,
      "commodity": 5001,
      "cpus": 6337,
      "accelerator": 807,
      "discrepancy": 8278,
      "lowering": 19292,
      "902": 593,
      "grace": 13182,
      "brains": 3852,
      "llama33b": 18225,
      "competitor": 5235,
      "struggling": 31250,
      "cheap": 4509,
      "aroused": 2508,
      "milestone": 20161,
      "expenditure": 10354,
      "tailoring": 32020,
      "pathology": 24399,
      "bolster": 3807,
      "indicated": 14698,
      "windows": 35615,
      "2048": 243,
      "verifies": 35215,
      "languageguided": 16859,
      "interpreter": 15466,
      "weeks": 35491,
      "pandagpt": 23998,
      "auditory": 2787,
      "audios": 2782,
      "scalar": 29127,
      "force": 11716,
      "naturalistic": 22582,
      "nonsequential": 23008,
      "backpack": 3064,
      "intervene": 15475,
      "6bparameter": 492,
      "imagine": 14097,
      "technically": 32613,
      "codedavinci002": 4837,
      "carriers": 4188,
      "determination": 7895,
      "assigning": 2625,
      "decoupling": 7295,
      "blend": 3760,
      "decides": 7228,
      "humancreated": 13895,
      "obvious": 23266,
      "deploys": 7662,
      "gqa": 13181,
      "biasing": 3689,
      "styled": 31415,
      "journey": 15718,
      "qformer": 26921,
      "intersection": 15470,
      "domainadaptive": 8606,
      "selfgenerated": 29420,
      "400": 375,
      "231": 257,
      "689": 487,
      "poorer": 25105,
      "unnecessary": 34417,
      "preprocessed": 25503,
      "inputted": 15052,
      "falling": 10954,
      "superni": 31667,
      "multi": 22264,
      "rouge1": 28981,
      "lasted": 17397,
      "underperforms": 34167,
      "patch": 24393,
      "affirmative": 1546,
      "composes": 5322,
      "deployments": 7661,
      "similarsized": 30137,
      "svd": 31788,
      "proactive": 25964,
      "amr": 1898,
      "caveat": 4259,
      "pick": 24945,
      "animal": 2011,
      "unlocks": 34414,
      "prp": 26801,
      "50x": 429,
      "uncertain": 34114,
      "instantiated": 15116,
      "humanagent": 13881,
      "let": 17725,
      "69": 488,
      "nonexpert": 22997,
      "seamless": 29302,
      "acoustic": 1113,
      "avaliable": 3003,
      "17b": 181,
      "negotiations": 22672,
      "unsuitable": 34452,
      "monitoring": 22207,
      "enlarge": 9580,
      "2chat": 295,
      "derivation": 7669,
      "generalisation": 12195,
      "equation": 9676,
      "humanunderstandable": 13935,
      "pathway": 24403,
      "80gb": 558,
      "800": 555,
      "noted": 23035,
      "bestfinetuned": 3581,
      "racial": 27143,
      "repair": 28086,
      "misunderstanding": 20243,
      "erroneous": 9706,
      "repairs": 28087,
      "reactions": 27292,
      "9b": 611,
      "indicator": 14706,
      "loglinear": 19163,
      "rejection": 27850,
      "359": 351,
      "scans": 29185,
      "inquiry": 15055,
      "implements": 14170,
      "encapsulates": 9334,
      "recorded": 27663,
      "california": 3981,
      "san": 29100,
      "exactmatch": 10093,
      "separates": 29585,
      "copilot": 6142,
      "544": 444,
      "voluminous": 35389,
      "transmission": 33860,
      "lexglue": 17797,
      "modals": 20328,
      "signifying": 30090,
      "transcending": 33662,
      "higherorder": 13615,
      "023": 6,
      "screening": 29293,
      "guanaco": 13328,
      "155": 146,
      "molecule": 22201,
      "httpsgithubcomnlpxucanwizardlm": 13775,
      "instructblip": 15132,
      "mme": 20310,
      "wanjuan": 35418,
      "confidential": 5653,
      "internlm": 15449,
      "closedsourced": 4691,
      "democratization": 7423,
      "humancentric": 13891,
      "llama27bchat": 18214,
      "synthesizes": 31842,
      "embracing": 9153,
      "compliance": 5305,
      "displayed": 8316,
      "noting": 23044,
      "lvlms": 19335,
      "qwenvlchat": 27139,
      "experienced": 10372,
      "2k": 297,
      "safetycritical": 29055,
      "anomaly": 2037,
      "lvlm": 19332,
      "anomalous": 2036,
      "vlm": 35376,
      "activity": 1152,
      "feeds": 11079,
      "dissemination": 8324,
      "clouds": 4704,
      "desiderata": 7694,
      "file": 11162,
      "attentionfree": 2748,
      "traininginference": 33651,
      "tuningfree": 34021,
      "neighborhood": 22673,
      "fkgl": 11582,
      "perlayer": 24861,
      "defect": 7357,
      "odds": 23280,
      "pre": 25384,
      "pareto": 24309,
      "understandability": 34204,
      "untested": 34461,
      "transitioning": 33809,
      "embark": 9123,
      "troubleshooting": 33922,
      "stop": 31073,
      "brand": 3855,
      "audiocaps": 2781,
      "audioset": 2783,
      "notorious": 23046,
      "asymmetric": 2675,
      "threestage": 33094,
      "llama70b": 18229,
      "contextrich": 5934,
      "renowned": 28082,
      "664": 480,
      "194": 192,
      "823": 566,
      "dozen": 8703,
      "usable": 34500,
      "678": 483,
      "attempting": 2707,
      "reductions": 27765,
      "generalise": 12196,
      "segmentlevel": 29372,
      "recognise": 27635,
      "ldm": 17460,
      "misalignment": 20209,
      "hacking": 13364,
      "mmhalbench": 20311,
      "penalizing": 24447,
      "llavabench": 18254,
      "40000": 376,
      "anthropic": 2088,
      "sensor": 29524,
      "influenza": 14847,
      "vaccine": 34985,
      "vaccines": 34986,
      "instrumental": 15310,
      "infectious": 14756,
      "tooluse": 33276,
      "mpt30b": 22247,
      "graphics": 13237,
      "primitives": 25929,
      "52": 433,
      "1d": 201,
      "06": 12,
      "multiscale": 22438,
      "urdu": 34491,
      "4635": 400,
      "informationdense": 14928,
      "arrive": 2511,
      "suit": 31593,
      "conflicting": 5665,
      "170": 170,
      "105": 59,
      "inferential": 14829,
      "subtracting": 31500,
      "wellstructured": 35529,
      "langchain": 16031,
      "responsive": 28524,
      "heart": 13490,
      "harnessed": 13455,
      "redefines": 27693,
      "influencing": 14845,
      "oversight": 23953,
      "warm": 35421,
      "enduser": 9444,
      "lemur": 17702,
      "observable": 23216,
      "durations": 8754,
      "forming": 11765,
      "commendable": 4977,
      "histories": 13728,
      "stitching": 31067,
      "resemblance": 28382,
      "accept": 809,
      "95k": 604,
      "alleviation": 1796,
      "ed": 8817,
      "multistage": 22440,
      "obviating": 23264,
      "altered": 1844,
      "datatypes": 7196,
      "arising": 2498,
      "rotations": 28976,
      "uncharted": 34122,
      "doc": 8499,
      "touvron": 33303,
      "10times": 64,
      "differentiate": 8164,
      "revision": 28836,
      "baichuan2": 3076,
      "straightforwardly": 31094,
      "sole": 30460,
      "3gb": 369,
      "wake": 35408,
      "depression": 7663,
      "recommending": 27656,
      "92": 596,
      "satisfied": 29108,
      "assumes": 2661,
      "underline": 34151,
      "striving": 31157,
      "scrutiny": 29298,
      "understudy": 34283,
      "defacto": 7355,
      "a10080gb": 619,
      "theorem": 33046,
      "proving": 26787,
      "385": 360,
      "git": 12737,
      "selfreflection": 29430,
      "provision": 26788,
      "decouples": 7294,
      "crowdworkers": 6435,
      "constituent": 5773,
      "emphasized": 9211,
      "embeds": 9149,
      "resembling": 28385,
      "element": 9098,
      "substantiates": 31489,
      "affirms": 1547,
      "communitybased": 5062,
      "enforce": 9450,
      "failing": 10910,
      "prover": 26678,
      "155b": 147,
      "overrely": 23950,
      "vq": 35393,
      "catching": 4219,
      "complements": 5247,
      "baby": 3050,
      "babylm": 3051,
      "zephyr7b": 35933,
      "tutorials": 34033,
      "possesses": 25203,
      "postediting": 25224,
      "engages": 9455,
      "081": 17,
      "gpt4vision": 13138,
      "cold": 4899,
      "facets": 10832,
      "skypile": 30319,
      "condensing": 5565,
      "strategically": 31098,
      "sourcing": 30576,
      "syntactical": 31828,
      "neighborhoods": 22674,
      "buffer": 3903,
      "22x": 255,
      "twopart": 34039,
      "sifting": 29941,
      "cogvlm": 4883,
      "exacerbates": 10086,
      "impacting": 14145,
      "invaluable": 15563,
      "fingpt": 11559,
      "176": 177,
      "16000": 159,
      "openorca": 23476,
      "cascade": 4191,
      "4677": 401,
      "selfverification": 29444,
      "speechbased": 30791,
      "v15": 34983,
      "resilience": 28394,
      "elimination": 9111,
      "irish": 15642,
      "maximise": 19701,
      "unfair": 34304,
      "percent": 24458,
      "arisen": 2496,
      "lowentropy": 19281,
      "leaks": 17501,
      "longlora": 19210,
      "verbal": 35203,
      "pal": 23987,
      "onestage": 23353,
      "heterogeneity": 13529,
      "186": 189,
      "package": 23962,
      "videobased": 35267,
      "pursuit": 26888,
      "depicting": 7628,
      "revised": 28835,
      "659": 475,
      "uncertaintyaware": 34120,
      "socially": 30435,
      "layered": 17434,
      "synchronized": 31805,
      "preferring": 25477,
      "nvidias": 23197,
      "megatronlm": 19787,
      "condensed": 5563,
      "prefixbased": 25481,
      "imaging": 14098,
      "flickr30k": 11619,
      "patches": 24394,
      "movements": 22236,
      "instagram": 15107,
      "tough": 33302,
      "chainofthoughts": 4305,
      "exorbitant": 10333,
      "sharp": 29792,
      "publiclyreleased": 26868,
      "audited": 2785,
      "accentuates": 808,
      "capitalizes": 4131,
      "projector": 26254,
      "800k": 557,
      "factbased": 10857,
      "acknowledges": 1112,
      "ineffectiveness": 14748,
      "molecular": 22200,
      "bells": 3344,
      "whistles": 35539,
      "multiimage": 22289,
      "kl": 15802,
      "ema": 9120,
      "pruner": 26806,
      "smartphone": 30408,
      "pope": 25110,
      "consume": 5815,
      "missions": 20224,
      "supervisor": 31700,
      "pull": 26876,
      "7b13b": 543,
      "975": 607,
      "23m": 260,
      "collects": 4938,
      "circumventing": 4572,
      "inaccuracy": 14433,
      "codellms": 4842,
      "august": 2830,
      "laid": 16018,
      "7bs": 548,
      "mistrals": 20242,
      "mistake": 20225,
      "flowbased": 11628,
      "mild": 20160,
      "older": 23335,
      "icd": 13972,
      "partitioned": 24360,
      "lacked": 16010,
      "capitalize": 4130,
      "cut": 6568,
      "worthwhile": 35844,
      "frontal": 11930,
      "cnns": 4713,
      "ethnicity": 9809,
      "accomplishments": 844,
      "181": 187,
      "relearning": 27897,
      "propelled": 26466,
      "clients": 4653,
      "highrisk": 13709,
      "partitioning": 24361,
      "merges": 19850,
      "inaugural": 14440,
      "summarisation": 31607,
      "recording": 27664,
      "evidencebased": 10068,
      "resistance": 28396,
      "guideline": 13355,
      "satisfactorily": 29106,
      "rendered": 28079,
      "2023a": 238,
      "xu": 35883,
      "2024": 240,
      "impossibility": 14224,
      "llama2s": 18223,
      "demographics": 7431,
      "238": 259,
      "underutilize": 34287,
      "deducing": 7307,
      "closedform": 4685,
      "450": 394,
      "assets": 2621,
      "vicunas": 35261,
      "trainingtime": 33652,
      "safetyaligned": 29054,
      "harmless": 13447,
      "attack": 2685,
      "attacks": 2690,
      "origin": 23697,
      "disrupts": 8322,
      "gpt35turbo16k": 13047,
      "faults": 11007,
      "trendy": 33891,
      "591": 454,
      "red": 27689,
      "teaming": 32599,
      "jailbreaking": 15698,
      "redteaming": 27694,
      "sec": 29317,
      "filings": 11165,
      "finqa": 11563,
      "prevention": 25862,
      "llava157b": 18252,
      "talent": 32040,
      "atom": 2679,
      "multilinguality": 22335,
      "tsne": 33950,
      "emit": 9200,
      "apibank": 2104,
      "7k": 549,
      "welldesigned": 35521,
      "confusing": 5674,
      "copes": 6140,
      "correctional": 6200,
      "residuals": 28393,
      "illustrations": 14052,
      "tripadvisor": 33912,
      "2s": 301,
      "416": 385,
      "safeguard": 29034,
      "8000": 556,
      "epoch": 9671,
      "082": 18,
      "020": 5,
      "color": 4943,
      "unforeseen": 34311,
      "tone": 33254,
      "nationality": 22497,
      "depressive": 7665,
      "experiencing": 10375,
      "blends": 3763,
      "256k": 275,
      "64k": 473,
      "128k": 102,
      "pressure": 25618,
      "unleash": 34388,
      "disadvantaged": 8248,
      "reimagines": 27834,
      "augmenter": 2821,
      "shortcoming": 29820,
      "objectlevel": 23213,
      "fueled": 11943,
      "couple": 6308,
      "shaped": 29776,
      "llava7b": 18253,
      "overflow": 23933,
      "posting": 25227,
      "replacements": 28103,
      "overlooks": 23949,
      "distributionally": 8401,
      "minority": 20205,
      "grand": 13211,
      "indications": 14705,
      "modulation": 22191,
      "368": 356,
      "755": 517,
      "coefficients": 4871,
      "permissively": 24865,
      "licensed": 17815,
      "507": 426,
      "6k": 493,
      "granular": 13215,
      "decay": 7220,
      "exchanges": 10182,
      "refinements": 27783,
      "corrected": 6196,
      "greedily": 13279,
      "540": 442,
      "enumeration": 9664,
      "billionscale": 3729,
      "fool": 11710,
      "immune": 14118,
      "850": 576,
      "nonexistent": 22995,
      "llavarlhf": 18255,
      "632": 470,
      "8times": 585,
      "visiolinguistic": 35286,
      "pioneers": 24961,
      "onpar": 23372,
      "promptings": 26401,
      "minigptv2": 20178,
      "historically": 13727,
      "deployable": 7631,
      "automates": 2874,
      "hallucinationfree": 13385,
      "llama27bbased": 18212,
      "481": 405,
      "mysterious": 22475,
      "discipline": 8258,
      "checked": 4515,
      "spent": 30806,
      "nonstationary": 23010,
      "53x": 440,
      "8192": 564,
      "812": 561,
      "144": 138,
      "hardly": 13426,
      "substructures": 31495,
      "221": 253,
      "undoes": 34296,
      "predominance": 25459,
      "recruit": 27672,
      "braininspired": 3851,
      "parietal": 24311,
      "privately": 25959,
      "eyes": 10811,
      "domainrelated": 8609,
      "formatted": 11759,
      "warrants": 35424,
      "coping": 6143,
      "yi": 35905,
      "977": 608,
      "saturation": 29112,
      "142": 136,
      "presently": 25571,
      "001": 0,
      "mamba": 19525,
      "carries": 4189,
      "overhaul": 23934,
      "sparser": 30624,
      "endowed": 9427,
      "partitions": 24362,
      "promises": 26279,
      "vllms": 35375,
      "movement": 22235,
      "rejected": 27849,
      "2023b": 239,
      "40x": 382,
      "misinterpret": 20214,
      "topology": 33295,
      "induces": 14733,
      "regulations": 27832,
      "circumvent": 4571,
      "distinguishes": 8376,
      "4bit": 412,
      "rectifying": 27675,
      "determines": 7901,
      "tale": 32039,
      "expenses": 10356,
      "divides": 8490,
      "condenses": 5564,
      "han": 13402,
      "severity": 29755,
      "entangled": 9618,
      "cos": 6238,
      "instructfollowing": 15135,
      "coordinated": 6136,
      "readiness": 27303,
      "inthewild": 15479,
      "spectrogram": 30770,
      "extraordinarily": 10784,
      "freeze": 11917,
      "prioritization": 25945,
      "clustered": 4707,
      "handles": 13412,
      "concurrent": 5560,
      "dependability": 7614,
      "errorfree": 9718,
      "timely": 33155,
      "comply": 5307,
      "thousand": 33084,
      "coordinate": 6135,
      "systems data": 31892,
      "crucial challenge": 6438,
      "models expected": 21233,
      "compute time": 5498,
      "mainly natural": 19405,
      "efficacy pretrained": 8989,
      "generation developed": 12485,
      "conducted extensive": 5635,
      "translation text": 33856,
      "gpt2 demonstrated": 12882,
      "ability capture": 658,
      "multiturn dialogue": 22463,
      "ngram analysis": 22901,
      "article describes": 2521,
      "using transformerbased": 34933,
      "models automated": 20997,
      "model retrained": 20758,
      "public domain": 26836,
      "contextualized representations": 5962,
      "produced models": 26158,
      "upper layers": 34483,
      "representations layers": 28165,
      "embedding word": 9135,
      "techniques shown": 32664,
      "annotated dataset": 2018,
      "processing transformer": 26132,
      "selfsupervised pretraining": 29440,
      "reduce required": 27724,
      "required number": 28230,
      "factor 10": 10864,
      "powerful tool": 25353,
      "labeled samples": 15957,
      "make better": 19453,
      "better quality": 3620,
      "quality text": 26981,
      "model openai": 20663,
      "automatic metrics": 2889,
      "text using": 32962,
      "decoding algorithms": 7270,
      "domain data": 8559,
      "data new": 6784,
      "correctness generated": 6208,
      "require extensive": 28214,
      "downstream nlp": 8682,
      "tasks pretrained": 32454,
      "generation exploration": 12500,
      "exploration paper": 10565,
      "model requires": 20754,
      "par stateoftheart": 24153,
      "generation aims": 12455,
      "rules generate": 29011,
      "transformer network": 33735,
      "model consisting": 20439,
      "model transformer": 20842,
      "input representation": 15023,
      "method produce": 19958,
      "performance proposed": 24724,
      "generation guided": 12514,
      "order generate": 23674,
      "range applications": 27187,
      "respectively leveraging": 28461,
      "knowledge embedded": 15839,
      "making use": 19517,
      "unsupervised learning": 34455,
      "finetuning present": 11485,
      "technique using": 32624,
      "sentence level": 29536,
      "spans text": 30602,
      "text smaller": 32944,
      "pairs isolating": 23980,
      "data automatically": 6616,
      "language text": 16834,
      "text english": 32852,
      "github repository": 12739,
      "paper devise": 24036,
      "require new": 28223,
      "distribution language": 8394,
      "given proposed": 12763,
      "capable producing": 4117,
      "parameter gpt2": 24182,
      "synthetic corpus": 31849,
      "83 billion": 569,
      "access models": 824,
      "apply methodology": 2209,
      "key problem": 15782,
      "works best": 35810,
      "massive training": 19633,
      "realworld scenario": 27346,
      "new domain": 22794,
      "domain ability": 8551,
      "target data": 32047,
      "standard method": 30880,
      "visual question": 35349,
      "remains challenge": 27992,
      "sound reasoning": 30543,
      "proposing new": 26632,
      "model generating": 20548,
      "visual questions": 35352,
      "weights using": 35518,
      "task predicting": 32178,
      "vqa generating": 35397,
      "tokens text": 33248,
      "generation gpt2": 12513,
      "specifically pretrained": 30753,
      "character level": 4422,
      "paper use": 24146,
      "recently introduced": 27604,
      "closer human": 4696,
      "human text": 13870,
      "model propose": 20732,
      "jensenshannon divergence": 15704,
      "generation understanding": 12627,
      "results wide": 28711,
      "modeling benchmarks": 20888,
      "deep generative": 7319,
      "era largescale": 9701,
      "largescale pretraining": 17379,
      "pretraining make": 25819,
      "methods practical": 20075,
      "emergence largescale": 9173,
      "remain unexplored": 27988,
      "aligned human": 1737,
      "highlight importance": 13630,
      "performance introduce": 24636,
      "better par": 3615,
      "requires model": 28257,
      "model track": 20831,
      "writing styles": 35858,
      "demonstrate largescale": 7469,
      "gpt2 grover": 12906,
      "gpt2 achieved": 12869,
      "constraints address": 5788,
      "simple novel": 30159,
      "generation proposed": 12580,
      "tokens existing": 33225,
      "finetune downstream": 11282,
      "time complexity": 33112,
      "simple language": 30155,
      "user input": 34653,
      "leads stateoftheart": 17496,
      "main metrics": 19399,
      "corresponding input": 6228,
      "text recent": 32929,
      "rely additional": 27969,
      "transformerbased unidirectional": 33770,
      "unidirectional language": 34321,
      "additional taskspecific": 1265,
      "introduced new": 15540,
      "race dataset": 27142,
      "users paper": 34696,
      "responses work": 28517,
      "procedure obtain": 26045,
      "demonstrate scaling": 7491,
      "yielded similar": 35917,
      "using bert": 34740,
      "covid19 pandemic": 6333,
      "learning approaches": 17538,
      "bert openai": 3521,
      "openai gpt2": 23435,
      "solve challenge": 30488,
      "evaluate results": 9863,
      "comprehensive information": 5380,
      "extracted original": 10750,
      "generate paraphrases": 12311,
      "classification experiments": 4596,
      "paraphrases generated": 24306,
      "good quality": 12823,
      "used data": 34594,
      "online communities": 23362,
      "media provide": 19762,
      "provide potential": 26721,
      "models possible": 21825,
      "model automatically": 20381,
      "model applied": 20371,
      "relevant sentences": 27945,
      "performance evaluation": 24585,
      "experts rate": 10521,
      "additionally based": 1273,
      "web application": 35473,
      "data generate": 6720,
      "generate natural": 12304,
      "power pretrained": 25328,
      "features different": 11031,
      "visual textual": 35357,
      "framework allows": 11828,
      "models capture": 21045,
      "potential direction": 25253,
      "generation long": 12542,
      "models largescale": 21432,
      "gpt2 powerful": 12935,
      "coherent long": 4893,
      "text various": 32963,
      "design allows": 7696,
      "set examples": 29685,
      "broad set": 3886,
      "methods terms": 20102,
      "quality sample": 26977,
      "lstm gpt2": 19327,
      "provide large": 26713,
      "attributes using": 2764,
      "trained classify": 33385,
      "weight distribution": 35494,
      "generated models": 12375,
      "achieved best": 1001,
      "times gpt2": 33161,
      "models draw": 21174,
      "achieve near": 972,
      "corpus used": 6188,
      "gpt 20": 12841,
      "zero shot": 35945,
      "work applies": 35669,
      "generate plausible": 12313,
      "generates valid": 12406,
      "previously unseen": 25904,
      "gpt2 finetuning": 12892,
      "offers novel": 23311,
      "vast data": 35184,
      "advantage using": 1501,
      "released models": 27926,
      "pretraining strategies": 25841,
      "respectively extensive": 28460,
      "analysis identify": 1926,
      "effective technique": 8898,
      "finetuning steps": 11538,
      "examples paper": 10139,
      "tasks importantly": 32359,
      "highlight current": 13628,
      "sentence sentence": 29543,
      "text human": 32891,
      "require manual": 28221,
      "quality existing": 26958,
      "reducing gap": 27747,
      "despite widespread": 7828,
      "investigate use": 15598,
      "hardware single": 13432,
      "synthetic text": 31861,
      "limited success": 17968,
      "generation contextual": 12479,
      "increasingly popular": 14641,
      "models prone": 21864,
      "reasonable perplexity": 27361,
      "easily identified": 8798,
      "improve coherence": 14259,
      "coherence consistency": 4887,
      "given topic": 12778,
      "outperformed baselines": 23792,
      "outperforms conventional": 23815,
      "models humanlike": 21365,
      "scoring model": 29286,
      "generation important": 12518,
      "problem lies": 25994,
      "reasoning module": 27424,
      "human writing": 13879,
      "generation automatic": 12461,
      "evaluation proposed": 9997,
      "models design": 21142,
      "level language": 17733,
      "following concept": 11688,
      "implementation perspective": 14163,
      "model transfer": 20841,
      "bestperforming models": 3585,
      "demonstrate possible": 7479,
      "visual input": 35334,
      "visual information": 35333,
      "successfully generates": 31539,
      "generation high": 12515,
      "variety settings": 35070,
      "new framework": 22805,
      "effort required": 9080,
      "augmentation finetuning": 2798,
      "finetuning text": 11548,
      "yelp reviews": 35902,
      "text including": 32896,
      "fluency experiments": 11631,
      "use dataset": 34523,
      "dataset existing": 6985,
      "preferences results": 25475,
      "architectures gpt2": 2463,
      "recurrent architectures": 27678,
      "parameter training": 24201,
      "generation generative": 12511,
      "tremendous success": 33884,
      "exhibit better": 10211,
      "result better": 28541,
      "generation particular": 12569,
      "decoding hyperparameters": 7273,
      "results multiple": 28648,
      "tasks especially": 32315,
      "multiple sources": 22420,
      "work addressed": 35665,
      "entire document": 9624,
      "model task": 20825,
      "twostage generation": 34044,
      "metrics human": 20139,
      "generate semantically": 12321,
      "active research": 1146,
      "topic generating": 33283,
      "work train": 35795,
      "model answer": 20366,
      "earlier work": 8772,
      "larger base": 17315,
      "models lead": 21437,
      "lead better": 17462,
      "agent generate": 1561,
      "accomplish tasks": 842,
      "uses pretrained": 34713,
      "corresponding instructions": 6229,
      "simulated data": 30187,
      "easier understand": 8789,
      "fully automated": 11953,
      "automated approaches": 2855,
      "approaches used": 2400,
      "information accurately": 14853,
      "new parallel": 22829,
      "roberta xlnet": 28924,
      "absolute improvement": 764,
      "construct knowledge": 5800,
      "manner requiring": 19549,
      "knowledge recent": 15900,
      "recent deep": 27512,
      "writing code": 35851,
      "propose unsupervised": 26582,
      "finetuning corpora": 11382,
      "new existing": 22802,
      "topics large": 33291,
      "framework analyzing": 11829,
      "techniques demonstrate": 32633,
      "framework conduct": 11837,
      "factors contribute": 10869,
      "biases models": 3680,
      "results confirm": 28583,
      "ranking models": 27234,
      "higher sensitivity": 13605,
      "sensitivity word": 29522,
      "model characteristics": 20414,
      "transformersbased models": 33800,
      "glove embeddings": 12800,
      "experiments benchmarks": 10422,
      "compared current": 5129,
      "editing approach": 8826,
      "leveraging abilities": 17777,
      "output model": 23875,
      "opens possibility": 23481,
      "grammatical correctness": 13205,
      "model provide": 20734,
      "comparison results": 5199,
      "field neural": 11148,
      "architectures models": 2468,
      "architecture used": 2456,
      "gpt2 order": 12932,
      "architecture gpt2": 2444,
      "architecture designed": 2441,
      "information present": 14898,
      "language captions": 16048,
      "adapt language": 1164,
      "caption generation": 4134,
      "language encoding": 16068,
      "essential role": 9760,
      "problems despite": 26025,
      "recent methods": 27534,
      "dataset paper": 7018,
      "presents novel": 25588,
      "technique named": 32623,
      "outperforms competitive": 23814,
      "introduce technique": 15537,
      "preserving semantic": 25612,
      "level specifically": 17736,
      "training sequence": 33608,
      "real life": 27311,
      "domains limited": 8630,
      "control content": 6048,
      "open questions": 23417,
      "large training": 17279,
      "based finetuning": 3161,
      "document retrieval": 8506,
      "expand users": 10336,
      "use text": 34574,
      "different experiments": 8077,
      "easily implemented": 8799,
      "gpt code": 12845,
      "produce highquality": 26148,
      "gpt2 accounts": 12868,
      "search space": 29312,
      "improvements model": 14358,
      "problem given": 25990,
      "obtain comparable": 23249,
      "new perspective": 22830,
      "generation effectiveness": 12490,
      "built pretrained": 3936,
      "gpt2 specifically": 12953,
      "health support": 13486,
      "reduce global": 27710,
      "key component": 15757,
      "studies shown": 31283,
      "shown highly": 29883,
      "rl agent": 28899,
      "network based": 22686,
      "evaluation demonstrate": 9938,
      "generation work": 12639,
      "generate large": 12299,
      "model small": 20801,
      "synthetic useful": 31862,
      "space search": 30583,
      "framework generate": 11860,
      "based clip": 3144,
      "image input": 14070,
      "language decoder": 16059,
      "architecture language": 2445,
      "conditional text": 5571,
      "learn generate": 17507,
      "generate labels": 12297,
      "achieving similar": 1103,
      "small training": 30371,
      "automatically annotated": 2902,
      "approach automatically": 2240,
      "supervised training": 31692,
      "training paradigm": 33583,
      "lack training": 16006,
      "text samples": 32936,
      "samples available": 29073,
      "novel fewshot": 23079,
      "samples data": 29075,
      "text text": 32956,
      "pretraining paper": 25830,
      "endtoend task": 9442,
      "evaluated human": 9877,
      "different traditional": 8151,
      "propose endtoend": 26508,
      "uses generative": 34710,
      "outperforms baselines": 23808,
      "lack knowledge": 15995,
      "dataset carefully": 6946,
      "automatic methods": 2887,
      "human written": 13880,
      "using automated": 34735,
      "numerous tasks": 23189,
      "provided natural": 26741,
      "models predicting": 21831,
      "models bias": 21026,
      "long document": 19170,
      "document summarization": 8507,
      "resource setting": 28416,
      "industrial settings": 14742,
      "document length": 8504,
      "compress long": 5402,
      "algorithm based": 1702,
      "perplexity scores": 24873,
      "baselines furthermore": 3266,
      "human labeling": 13836,
      "syntax trees": 31834,
      "performances various": 24822,
      "syntactic information": 31817,
      "information syntactic": 14914,
      "crucial success": 6451,
      "effectively efficiently": 8915,
      "paper address": 24002,
      "problem proposing": 26006,
      "architecture experiments": 2443,
      "experiments various": 10498,
      "datasets natural": 7152,
      "achieve consistent": 958,
      "types pretraining": 34069,
      "architectures including": 2464,
      "tasks main": 32410,
      "main categories": 19389,
      "unconditional generation": 34131,
      "generation conditional": 12475,
      "positional encodings": 25188,
      "varying number": 35177,
      "conditional unconditional": 5573,
      "given model": 12755,
      "sizes data": 30297,
      "bert large": 3516,
      "context generation": 5892,
      "successful application": 31532,
      "boost accuracy": 3815,
      "responsible extracting": 28523,
      "networks way": 22718,
      "showing proposed": 29864,
      "representations learned": 28166,
      "stateoftheart transformerbased": 31004,
      "context training": 5923,
      "novel effective": 23074,
      "bert variants": 3536,
      "results zeroshot": 28714,
      "computer pc": 5503,
      "prompts condition": 26408,
      "prompts used": 26446,
      "model tuning": 20844,
      "weights tuned": 35516,
      "model multiple": 20649,
      "attention based": 2713,
      "toxic nontoxic": 33311,
      "gpt2 glove": 12899,
      "results experimental": 28608,
      "span tokens": 30592,
      "benchmark natural": 3402,
      "visionlanguage tasks": 35324,
      "recently increasing": 27601,
      "methods lack": 20056,
      "evaluation frameworks": 9953,
      "scarcity datasets": 29190,
      "introduce evil": 15505,
      "provides comprehensive": 26748,
      "approaches generate": 2374,
      "largest existing": 17392,
      "generation surpasses": 12609,
      "task identifying": 32137,
      "received attention": 27477,
      "attention language": 2720,
      "analyze capabilities": 1988,
      "educational settings": 8845,
      "offtheshelf language": 23327,
      "results raise": 28665,
      "adopt curriculum": 1401,
      "data simple": 6866,
      "set conditions": 29678,
      "learning procedure": 17633,
      "ai model": 1611,
      "uses gpt2": 34711,
      "model order": 20665,
      "recently models": 27611,
      "apply new": 2211,
      "language description": 16060,
      "poses new": 25171,
      "stateoftheart vision": 31006,
      "models transformerbased": 22083,
      "endtoend manner": 9436,
      "work qualitative": 35770,
      "experiments verify": 10501,
      "work pave": 35742,
      "way future": 35433,
      "meaning neural": 19719,
      "effectiveness neural": 8962,
      "contextual word": 5956,
      "functional similarities": 11965,
      "dynamic semantics": 8763,
      "massively multilingual": 19635,
      "investigate impact": 15585,
      "modeling objectives": 20902,
      "making hard": 19504,
      "tasks 34": 32227,
      "come important": 4968,
      "highlighting need": 13646,
      "focus zeroshot": 11659,
      "pretraining tasks": 25845,
      "relative position": 27882,
      "cross entropy": 6405,
      "knowledge language": 15869,
      "multiple facts": 22393,
      "efficiency propose": 9011,
      "popular pretrained": 25136,
      "leads faster": 17493,
      "faster training": 11003,
      "training higher": 33527,
      "representation models": 28147,
      "models preserve": 21837,
      "contextual knowledge": 5951,
      "relationships input": 27875,
      "textual modality": 33033,
      "visual language": 35341,
      "modalities images": 20323,
      "process interpretability": 26065,
      "bias detection": 3643,
      "stateoftheart multimodal": 30960,
      "provide services": 26728,
      "openai gpt": 23434,
      "showed finetuned": 29856,
      "compared pretrained": 5161,
      "social bias": 30417,
      "social biases": 30418,
      "biases study": 3686,
      "assessing bias": 2609,
      "biases present": 3682,
      "questionanswering systems": 27088,
      "risks posed": 28895,
      "datasets like": 7143,
      "current popular": 6523,
      "suffer issues": 31550,
      "tackle problems": 32000,
      "text better": 32822,
      "capacity models": 4128,
      "generate potential": 12314,
      "model including": 20577,
      "finetuning representation": 11508,
      "benchmarks method": 3461,
      "log analysis": 19151,
      "recently received": 27618,
      "answer candidates": 2040,
      "approach does": 2261,
      "increasing performance": 14627,
      "annotated examples": 2021,
      "bert generative": 3505,
      "bert bidirectional": 3498,
      "language prior": 16773,
      "rate reduction": 27268,
      "documents leveraging": 8518,
      "addresses problem": 1367,
      "particular study": 24341,
      "systems including": 31901,
      "components natural": 5315,
      "critical component": 6384,
      "easily extendable": 8797,
      "toxic responses": 33313,
      "study controllable": 31311,
      "applying method": 2219,
      "problem pretrained": 26003,
      "unsatisfactory performance": 34431,
      "design novel": 7712,
      "entity generation": 9644,
      "generation experimental": 12496,
      "dataset demonstrate": 6969,
      "performance automatic": 24527,
      "current applications": 6481,
      "rl finetuning": 28902,
      "making available": 19497,
      "10 absolute": 25,
      "outside training": 23901,
      "gpt2 decoder": 12880,
      "capability generating": 4090,
      "similar gpt2": 30103,
      "modeling generation": 20894,
      "codes models": 4853,
      "existing linguistic": 10287,
      "models consistent": 21099,
      "systems recently": 31917,
      "inference model": 14793,
      "challenging previous": 4392,
      "conducted benchmark": 5629,
      "vital tool": 35372,
      "gpt2 work": 12968,
      "different modules": 8110,
      "tasks labeled": 32385,
      "plms paper": 25049,
      "raw input": 27283,
      "downstream natural": 8678,
      "promising performance": 26293,
      "performance lack": 24639,
      "communication people": 5052,
      "responses model": 28500,
      "generate invalid": 12296,
      "stateoftheart solutions": 30991,
      "generation lack": 12530,
      "deteriorates performance": 7892,
      "bert transformer": 3535,
      "groups based": 13304,
      "reasonably good": 27364,
      "deployment language": 7649,
      "evaluate multilingual": 9852,
      "multiclass classification": 22273,
      "models predict": 21830,
      "incontext fewshot": 14545,
      "results language": 28635,
      "specifically study": 30756,
      "examples include": 10126,
      "simpler tasks": 30168,
      "compare zeroshot": 5116,
      "tasks categories": 32258,
      "categories compared": 4222,
      "effective future": 8872,
      "produce fluent": 26145,
      "fluent text": 11638,
      "additional models": 1261,
      "approach works": 2360,
      "leads diverse": 17491,
      "user studies": 34674,
      "method leverages": 19942,
      "model time": 20830,
      "method directly": 19902,
      "recurrent model": 27680,
      "open book": 23384,
      "closed book": 4675,
      "given significant": 12771,
      "settings propose": 29742,
      "texts social": 33000,
      "based review": 3218,
      "t5 finetuned": 31945,
      "performance suggesting": 24772,
      "general nlp": 12180,
      "provides useful": 26766,
      "traditional nlp": 33350,
      "task training": 32202,
      "symbolic reasoning": 31803,
      "example training": 10111,
      "promising area": 26283,
      "area nlp": 2481,
      "nlp field": 22933,
      "model method": 20641,
      "domains compared": 8614,
      "furthermore evaluated": 11999,
      "evaluated proposed": 9882,
      "comprehensive ablation": 5351,
      "structured prediction": 31227,
      "form knowledge": 11739,
      "distillation kd": 8339,
      "teacher using": 32586,
      "different prompt": 8126,
      "hallucination rate": 13383,
      "significantly higher": 30054,
      "using examples": 34774,
      "requiring training": 28276,
      "inference systems": 14812,
      "paper revisit": 24131,
      "respond appropriately": 28467,
      "specifically finetuning": 30741,
      "finetuning propose": 11499,
      "coherent responses": 4894,
      "generation providing": 12581,
      "associated lower": 2650,
      "core idea": 6153,
      "novelty lies": 23127,
      "method approach": 19880,
      "improve accuracy": 14256,
      "leverage additional": 17743,
      "adding additional": 1232,
      "model predicts": 20717,
      "given input": 12749,
      "input image": 15012,
      "present simple": 25553,
      "approach address": 2232,
      "clip model": 4666,
      "contains rich": 5839,
      "model additional": 20357,
      "clip language": 4665,
      "quantitative evaluation": 26992,
      "meaning representation": 19721,
      "utilize pretrained": 34962,
      "performance response": 24740,
      "experiments response": 10477,
      "growth training": 13323,
      "community witnessed": 5061,
      "analysis widely": 1979,
      "adopted transformer": 1405,
      "xlnet electra": 35879,
      "superior results": 31657,
      "prompt based": 26308,
      "based method": 3194,
      "relatively fewer": 27886,
      "structural information": 31210,
      "baselines significant": 3269,
      "problem annotating": 25986,
      "presented task": 25568,
      "require costly": 28213,
      "generating valid": 12446,
      "approach jointly": 2304,
      "models benefit": 21014,
      "approach effectively": 2266,
      "data outperform": 6790,
      "substantial margin": 31469,
      "knowledge topic": 15913,
      "models assess": 20989,
      "tuning gpt2": 33980,
      "model parameterefficient": 20688,
      "industrial applications": 14741,
      "applications diverse": 2149,
      "adapt new": 1167,
      "overhead work": 23938,
      "parameters prime": 24279,
      "systems recent": 31916,
      "recent development": 27513,
      "achieving stateoftheart": 1106,
      "generate responses": 12320,
      "study effectiveness": 31322,
      "substantial performance": 31473,
      "gptbased models": 13141,
      "understanding paper": 34259,
      "model relatively": 20749,
      "robustness performance": 28949,
      "class similar": 4586,
      "results common": 28579,
      "including finetuned": 14476,
      "remarkable consistency": 28040,
      "performance adversarial": 24521,
      "image classifiers": 14063,
      "gender race": 12152,
      "people different": 24451,
      "images using": 14090,
      "model apply": 20372,
      "text finetuned": 32859,
      "model frozen": 20539,
      "accuracy raw": 906,
      "large size": 17277,
      "theory experiments": 33058,
      "token time": 33204,
      "bias machine": 3655,
      "text uses": 32961,
      "representation words": 28152,
      "tasks maintains": 32415,
      "predict masked": 25409,
      "masked tokens": 19618,
      "methods problem": 20079,
      "multiple words": 22428,
      "method paper": 19954,
      "hidden layer": 13536,
      "tokens time": 33249,
      "time explore": 33125,
      "time consumption": 33115,
      "application large": 2130,
      "high variance": 13589,
      "environments recent": 9669,
      "language finetuned": 16078,
      "domains results": 8637,
      "models hope": 21360,
      "modeling techniques": 20910,
      "tasks completely": 32270,
      "original prompt": 23719,
      "prompt model": 26336,
      "taskspecific model": 32565,
      "model output": 20676,
      "prompt outputs": 26339,
      "prompt models": 26337,
      "prompt continuous": 26316,
      "continuous vectors": 6000,
      "model models": 20645,
      "performance challenging": 24540,
      "challenging datasets": 4381,
      "datasets currently": 7088,
      "models image": 21369,
      "number trainable": 23167,
      "gpt2 vision": 12967,
      "vision encoder": 35299,
      "framework achieves": 11824,
      "baseline systems": 3259,
      "fewer trainable": 11092,
      "fewer data": 11086,
      "generation recently": 12592,
      "learning method": 17602,
      "scratch using": 29292,
      "model lstm": 20632,
      "synthesized dataset": 31841,
      "model orders": 20666,
      "provide useful": 26732,
      "evaluation experiments": 9945,
      "textual modalities": 33032,
      "training phase": 33586,
      "provides stateoftheart": 26763,
      "directly finetuning": 8237,
      "gpt2 generation": 12897,
      "set small": 29706,
      "methods achieve": 20000,
      "results especially": 28603,
      "representation produced": 28149,
      "gpt2 text": 12957,
      "method reduces": 19965,
      "example words": 10114,
      "early layer": 8777,
      "lexical word": 17804,
      "words semantically": 35661,
      "providing natural": 26776,
      "improving task": 14423,
      "extremely computationally": 10796,
      "models feasible": 21264,
      "approach improving": 2298,
      "humans automatically": 13920,
      "430 percentage": 390,
      "accuracy code": 864,
      "contextualizing language": 5966,
      "related work": 27859,
      "semantic properties": 29465,
      "adapts gpt2": 1225,
      "significantly mitigates": 30069,
      "visual representations": 35356,
      "study generative": 31336,
      "foster research": 11786,
      "make attempt": 19451,
      "comparison extractive": 5194,
      "short context": 29812,
      "outperforms standard": 23851,
      "study effect": 31321,
      "fail generalize": 10901,
      "pretrained natural": 25735,
      "highquality short": 13699,
      "longer texts": 19202,
      "time control": 33116,
      "text domain": 32848,
      "methods finetuning": 20039,
      "text structure": 32950,
      "text length": 32908,
      "analysis largescale": 1938,
      "predicting human": 25417,
      "despite advances": 7769,
      "largescale annotated": 17343,
      "dataset assess": 6940,
      "set best": 29672,
      "test generated": 32769,
      "models observe": 21763,
      "generations finetuned": 12643,
      "build generative": 3911,
      "increases model": 14614,
      "data lowresource": 6761,
      "time costs": 33118,
      "upstream data": 34487,
      "uses update": 34717,
      "straightforward effective": 31093,
      "method combines": 19890,
      "retrieval effective": 28740,
      "outperforms nonretrieval": 23838,
      "united states": 34368,
      "lack research": 16001,
      "standard dataset": 30872,
      "finetuning achieved": 11368,
      "generation building": 12465,
      "difficult train": 8174,
      "task lie": 32152,
      "effective ways": 8907,
      "constructed data": 5806,
      "ones experiments": 23347,
      "superiority method": 31664,
      "incurs significant": 14666,
      "time speedups": 33144,
      "decoderonly architecture": 7252,
      "results comparable": 28580,
      "efficient neural": 9051,
      "neural rankers": 22756,
      "models modern": 21737,
      "modern baselines": 22156,
      "vocabulary input": 35383,
      "produce correct": 26142,
      "networks bert": 22708,
      "bert embeddings": 3501,
      "art performance": 2519,
      "diverse generation": 8430,
      "robust adversarial": 28930,
      "adversarial samples": 1518,
      "language prompts": 16807,
      "predicted output": 25414,
      "performance settings": 24750,
      "lags far": 16017,
      "explore methods": 10592,
      "advantage fact": 1498,
      "multiple prompts": 22413,
      "data directly": 6678,
      "stateoftheart zeroshot": 31008,
      "accuracy gains": 883,
      "examples language": 10132,
      "generation remains": 12594,
      "modalities text": 20326,
      "semantically related": 29485,
      "zeroshot image": 35979,
      "visually grounded": 35367,
      "image text": 14076,
      "prompt sentence": 26342,
      "does introduce": 8532,
      "use evaluation": 34527,
      "presents unique": 25599,
      "unique challenge": 34355,
      "effectiveness context": 8940,
      "pretraining setup": 25837,
      "setup paper": 29749,
      "universally effective": 34374,
      "effective datasets": 8865,
      "present generalized": 25533,
      "unified perspective": 34335,
      "different pretraining": 8125,
      "downstream finetuning": 8677,
      "memory storage": 19830,
      "storage costs": 31076,
      "processing training": 26131,
      "methods offers": 20070,
      "offers alternative": 23306,
      "lower computational": 19285,
      "way introduce": 35438,
      "peft method": 24436,
      "stronger performance": 31201,
      "relatively tiny": 27895,
      "used experiments": 34600,
      "efficient trainingfree": 9062,
      "data significant": 6863,
      "study propose": 31381,
      "generation need": 12561,
      "need training": 22644,
      "process effectively": 26056,
      "applied gpt2": 2190,
      "baselines terms": 3272,
      "overall quality": 23911,
      "mainstream approach": 19409,
      "explore possibility": 10596,
      "model support": 20816,
      "reduce demand": 27708,
      "training separate": 33607,
      "model similar": 20785,
      "similar gpt3": 30104,
      "finetuning negligible": 11463,
      "employ techniques": 9255,
      "early exiting": 8775,
      "reduce inference": 27715,
      "size demonstrate": 30246,
      "personalized content": 24889,
      "content creation": 5855,
      "humanlevel performance": 13905,
      "spectrum natural": 30772,
      "largely unexplored": 17312,
      "text work": 32967,
      "directly training": 8244,
      "match score": 19646,
      "method advantage": 19874,
      "large transformer": 17280,
      "text image": 32894,
      "increasing size": 14630,
      "size language": 30253,
      "texttoimage models": 33007,
      "latent diffusion": 17406,
      "biases promptbased": 3684,
      "largescale multitask": 17366,
      "different forms": 8078,
      "unlike training": 34406,
      "data released": 6833,
      "behavior models": 3321,
      "internal states": 15442,
      "study present": 31376,
      "available dataset": 2971,
      "decoder models": 7248,
      "stateoftheart natural": 30962,
      "selects salient": 29402,
      "demonstrate lightweight": 7471,
      "evaluating robustness": 9914,
      "reasoning understanding": 27463,
      "understanding underlying": 34278,
      "experiments roberta": 10479,
      "showing models": 29860,
      "models robust": 21948,
      "evaluation sets": 10009,
      "generation context": 12478,
      "training context": 33457,
      "demonstrate substantial": 7502,
      "showing gains": 29859,
      "main challenge": 19390,
      "achieve coverage": 959,
      "empirically study": 9250,
      "neuron activation": 22766,
      "models integrating": 21402,
      "objective help": 23205,
      "accurately achieve": 931,
      "largescale neural": 17369,
      "models preference": 21835,
      "findings propose": 11245,
      "quality extensive": 26960,
      "learning words": 17670,
      "larger larger": 17325,
      "increase computational": 14593,
      "requirements recent": 28242,
      "recent models": 27535,
      "research proposing": 28350,
      "generation reranking": 12595,
      "reranking approach": 28279,
      "output large": 23869,
      "produce impressive": 26150,
      "scenarios requiring": 29218,
      "key information": 15772,
      "information long": 14885,
      "current models": 6516,
      "models offer": 21765,
      "variety domains": 35059,
      "novel way": 23125,
      "generating sentence": 12443,
      "high average": 13551,
      "work considers": 35681,
      "entire sentence": 9626,
      "generated sentence": 12388,
      "past year": 24391,
      "highlighting importance": 13645,
      "research opendomain": 28338,
      "retrieval module": 28749,
      "area natural": 2478,
      "model naturally": 20654,
      "lm perform": 19060,
      "operations recent": 23574,
      "ordinary differential": 23686,
      "differential equations": 8163,
      "relevant data": 27938,
      "cognitive overload": 4879,
      "method domain": 19905,
      "domain pretrained": 8583,
      "tackling problem": 32007,
      "zeroshot scenario": 35993,
      "rely heavily": 27970,
      "language construct": 16053,
      "randomly selected": 27183,
      "codeswitching datasets": 4859,
      "making possible": 19513,
      "reasonable time": 27362,
      "diverse dataset": 8423,
      "premises conclusions": 25495,
      "annotations automatically": 2030,
      "automatically constitute": 2905,
      "gptneox opt": 13161,
      "unlike image": 34395,
      "problem incorporating": 25991,
      "models clip": 21055,
      "generation incorporating": 12520,
      "generation conduct": 12476,
      "acquire general": 1116,
      "use different": 34525,
      "scores using": 29284,
      "clinical data": 4657,
      "sql generation": 30839,
      "predictive accuracy": 25448,
      "requires expensive": 28250,
      "generate valid": 12338,
      "performance benchmark": 24530,
      "generation finally": 12505,
      "lms large": 19093,
      "does hold": 8529,
      "125m 175b": 98,
      "lms provided": 19107,
      "lm types": 19065,
      "average score": 3024,
      "highlighting critical": 13643,
      "study prompt": 31380,
      "50 fewer": 421,
      "model lightweight": 20613,
      "designed test": 7746,
      "huge model": 13779,
      "modeling capabilities": 20889,
      "capabilities remains": 4067,
      "perform competitively": 24475,
      "generalization downstream": 12211,
      "unlabeled corpus": 34380,
      "language modelbased": 16213,
      "models opensourced": 21772,
      "supports various": 31719,
      "demo video": 7421,
      "llms memory": 18798,
      "sampling produces": 29095,
      "powerful paradigm": 25350,
      "knowledgeintensive nlp": 15930,
      "challenges particularly": 4366,
      "model offers": 20661,
      "consistently significantly": 5757,
      "largest chinese": 17391,
      "leverage unique": 17761,
      "post training": 25221,
      "performance loss": 24667,
      "models importantly": 21371,
      "2080 ti": 245,
      "code training": 4826,
      "training logs": 33556,
      "testing human": 32804,
      "achieve accuracy": 942,
      "trains lm": 33654,
      "lm generate": 19056,
      "lm trained": 19062,
      "14 tasks": 135,
      "20 average": 207,
      "received recent": 27483,
      "humanlevel accuracy": 13904,
      "attention paper": 2730,
      "manipulated adversarial": 19541,
      "reduce computation": 27701,
      "efficiency significantly": 9013,
      "generate test": 12329,
      "systematic evaluation": 31870,
      "energy consumption": 9447,
      "solving linear": 30509,
      "proven difficult": 26673,
      "works inference": 35814,
      "external sources": 10734,
      "using highquality": 34794,
      "web search": 35478,
      "demonstrate retrieval": 7490,
      "improve effectiveness": 14263,
      "clinical domain": 4658,
      "unseen languages": 34440,
      "japanese russian": 15700,
      "russian chinese": 29020,
      "carefully aligned": 4171,
      "finetuning different": 11389,
      "result shows": 28546,
      "languages pretraining": 16905,
      "medicine finance": 19772,
      "neural autoregressive": 22722,
      "models queries": 21882,
      "importance sampling": 14191,
      "generative architecture": 12651,
      "systems employ": 31894,
      "memory compute": 19807,
      "generation extensive": 12503,
      "capability llms": 4098,
      "llms popular": 18855,
      "fact verification": 10856,
      "research code": 28297,
      "challenging scarcity": 4397,
      "scarcity labeled": 29192,
      "highquality domain": 13687,
      "propose prompt": 26560,
      "prompts induce": 26424,
      "texts generated": 32992,
      "methods addition": 20003,
      "goal generate": 12806,
      "task existing": 32117,
      "lms code": 19075,
      "approach code": 2249,
      "lm codex": 19055,
      "nlp classification": 22924,
      "toxicity detection": 33319,
      "classification performs": 4603,
      "practical approach": 25362,
      "gained attention": 12060,
      "integrate goal": 15320,
      "evaluation human": 9959,
      "limited performance": 17959,
      "models texttoimage": 22059,
      "synthesis tasks": 31837,
      "scene graph": 29224,
      "operations extensive": 23570,
      "address shortcomings": 1357,
      "shortcomings propose": 29822,
      "strategies require": 31112,
      "t5 text": 31964,
      "text ranking": 32927,
      "ranking based": 27231,
      "limited studies": 17967,
      "classification rely": 4608,
      "achieve substantial": 994,
      "constrained decoding": 5780,
      "showing significant": 29865,
      "obtain consistent": 23250,
      "data underlying": 6899,
      "underlying difficulty": 34153,
      "report describes": 28112,
      "allows use": 1817,
      "use external": 34529,
      "single nvidia": 30217,
      "v100 gpu": 34982,
      "tuning prompt": 34009,
      "parameter efficiency": 24178,
      "data prompt": 6810,
      "tuning performs": 34005,
      "performs comparably": 24847,
      "samples fewshot": 29077,
      "fails match": 10915,
      "match performance": 19645,
      "approaches source": 2394,
      "contrastive search": 6024,
      "text autoregressive": 32819,
      "great importance": 13250,
      "model obtained": 20660,
      "autoregressive lms": 2951,
      "training study": 33626,
      "languages experimental": 16872,
      "set multimodal": 29697,
      "modeling image": 20896,
      "captioning visual": 4138,
      "storytelling speech": 31090,
      "comparable stateoftheart": 5091,
      "larger lms": 17327,
      "produces higher": 26161,
      "works model": 35818,
      "subjectverb agreement": 31429,
      "contextual representations": 5954,
      "report results": 28125,
      "manually crafted": 19564,
      "perform par": 24499,
      "diffusion language": 8181,
      "despite growing": 7782,
      "diffusionbased language": 8187,
      "model key": 20597,
      "blocks text": 3778,
      "generation benchmarks": 12464,
      "extra advantage": 10738,
      "natural solution": 22581,
      "underexplored work": 34144,
      "work conduct": 35679,
      "extensive studies": 10713,
      "task including": 32138,
      "model exceeds": 20497,
      "models apply": 20984,
      "various stateoftheart": 35147,
      "machinetranslated english": 19375,
      "language dataset": 16057,
      "models learning": 21442,
      "demonstration examples": 7595,
      "perform various": 24507,
      "learning examples": 17564,
      "acquire information": 1118,
      "propose reinforcement": 26561,
      "strong abilities": 31159,
      "models suggesting": 22029,
      "lack domain": 15984,
      "produces stateoftheart": 26163,
      "generate artificial": 12265,
      "results deep": 28585,
      "predictive performance": 25453,
      "lms lack": 19092,
      "ranging 1b": 27218,
      "176b parameters": 179,
      "parameters different": 24240,
      "content unfaithful": 5874,
      "faithfulness generated": 10931,
      "generated summaries": 12389,
      "datasets observe": 7156,
      "news datasets": 22879,
      "datasets given": 7126,
      "new metric": 22822,
      "model paradigm": 20685,
      "models considering": 21098,
      "limited use": 17973,
      "distilling reasoning": 8358,
      "approaches like": 2380,
      "effective inducing": 8875,
      "given new": 12758,
      "finally investigate": 11197,
      "models sentence": 21964,
      "evaluated accuracy": 9872,
      "accuracy identifying": 889,
      "methods demonstrate": 20019,
      "fail identify": 10903,
      "exhibit emergent": 10215,
      "ability learn": 694,
      "sentiment labels": 29570,
      "learning new": 17622,
      "fewshot evaluation": 11101,
      "models reliance": 21923,
      "models memory": 21715,
      "models sensitive": 21962,
      "turing test": 34026,
      "task challenges": 32091,
      "prompt chatgpt": 26311,
      "original generated": 23705,
      "specific inputs": 30696,
      "generally applied": 12240,
      "results approach": 28570,
      "performance bert": 24533,
      "roberta bart": 28915,
      "performance faster": 24596,
      "faster inference": 11000,
      "minimal modifications": 20190,
      "model analysis": 20365,
      "majority inference": 19448,
      "achieving state": 1104,
      "finetune student": 11301,
      "larger teacher": 17338,
      "arithmetic commonsense": 2500,
      "commonsense symbolic": 5044,
      "generation prompting": 12578,
      "applied text": 2199,
      "suffer low": 31552,
      "approach text": 2347,
      "data representations": 6836,
      "span multiple": 30591,
      "fewshot baselines": 11097,
      "transformers large": 33788,
      "llms subject": 18974,
      "significantly advancing": 30032,
      "advancing field": 1492,
      "bloom llms": 3787,
      "summarization text": 31627,
      "efforts focus": 9091,
      "syntax semantics": 31833,
      "field paper": 11149,
      "llms automated": 18439,
      "achieve goal": 963,
      "behavior terms": 3322,
      "length reduced": 17710,
      "varying complexities": 35169,
      "common nlp": 5010,
      "analysis indicates": 1928,
      "plms t5": 25050,
      "large publicly": 17272,
      "prohibitively large": 26242,
      "work apply": 35670,
      "effective improving": 8873,
      "models addition": 20958,
      "addition discover": 1238,
      "add new": 1228,
      "following task": 11703,
      "languages code": 16865,
      "speech target": 30789,
      "crossmodal representation": 6424,
      "representation alignment": 28138,
      "trained public": 33420,
      "matches performance": 19656,
      "despite using": 7825,
      "modelbased evaluation": 20880,
      "metrics text": 20148,
      "proposed evaluation": 26598,
      "built gpt2": 3933,
      "reliable evaluation": 27954,
      "increasing scale": 14629,
      "new qualitative": 22835,
      "gpt3 results": 13005,
      "design large": 7707,
      "large computation": 16935,
      "multitask settings": 22455,
      "scarcity long": 29194,
      "knowledge knowledge": 15868,
      "plan make": 24995,
      "attribute relation": 2758,
      "text snippets": 32945,
      "attributes types": 2763,
      "scaling language": 29165,
      "meta learning": 19858,
      "shown finetuning": 29877,
      "tasks described": 32293,
      "understanding performance": 34260,
      "scale diversity": 29132,
      "benchmark different": 3382,
      "different task": 8148,
      "finally finetuning": 11195,
      "paper characterize": 24020,
      "model benchmark": 20393,
      "end create": 9410,
      "task categories": 32090,
      "framework measure": 11882,
      "present insights": 25538,
      "different evaluation": 8075,
      "tasks input": 32374,
      "promptsource flan": 26450,
      "additional layer": 1258,
      "provided gpt2": 26739,
      "massive text": 19631,
      "readable text": 27297,
      "precision model": 25392,
      "years tasks": 35899,
      "finetune models": 11296,
      "models specific": 21998,
      "datasets necessary": 7154,
      "techniques paper": 32654,
      "retrieval models": 28748,
      "gptj 6b": 13146,
      "accuracy interpretability": 893,
      "models highly": 21358,
      "positions sequence": 25191,
      "sizes configurations": 30296,
      "observations propose": 23223,
      "matching visual": 19661,
      "text information": 32897,
      "motivated propose": 22229,
      "text propose": 32924,
      "task sequentially": 32192,
      "generative design": 12660,
      "compared proprietary": 5167,
      "prompt contrast": 26317,
      "average gain": 3018,
      "texts using": 33004,
      "automatic quantitative": 2892,
      "using templatebased": 34926,
      "create set": 6353,
      "similar large": 30106,
      "accuracy model": 898,
      "better training": 3632,
      "finetuned english": 11308,
      "provide analysis": 26684,
      "limited model": 17957,
      "specific target": 30717,
      "balance tradeoff": 3081,
      "scaling curve": 29161,
      "data format": 6717,
      "model checkpoint": 20416,
      "selection method": 29392,
      "semantic coherence": 29449,
      "learning recent": 17642,
      "capture abstract": 4144,
      "approach named": 2317,
      "datasets pretrained": 7161,
      "quality generation": 26965,
      "generation issue": 12528,
      "issue given": 15653,
      "zeroshot instruction": 35981,
      "ranging size": 27220,
      "quality assurance": 26942,
      "time budget": 33109,
      "stakeholders including": 30863,
      "t5 chatgpt": 31938,
      "emerging research": 9197,
      "empower users": 9266,
      "language focus": 16079,
      "position embeddings": 25182,
      "using constrained": 34756,
      "clinical language": 4659,
      "trained primarily": 33419,
      "safety critical": 29046,
      "critical domains": 6386,
      "smaller domainspecific": 30374,
      "domainspecific language": 8652,
      "experiments train": 10493,
      "data pretraining": 6805,
      "models match": 21702,
      "representative set": 28186,
      "configurations large": 5657,
      "set fewshot": 29687,
      "text explore": 32856,
      "text generative": 32887,
      "model assess": 20377,
      "assess generated": 2593,
      "nli task": 22913,
      "error types": 9717,
      "creative writing": 6371,
      "generation approaches": 12459,
      "amazon mechanical": 1869,
      "written humans": 35862,
      "possibilities future": 25205,
      "support systems": 31712,
      "generation prompt": 12577,
      "generate outputs": 12310,
      "mitigation strategies": 20267,
      "answering vqa": 2081,
      "models visual": 22118,
      "content particular": 5868,
      "evaluating multilingual": 9908,
      "used benchmark": 34586,
      "present details": 25524,
      "vit pretrained": 35369,
      "pretrained vision": 25775,
      "researchers explore": 28376,
      "lms increasingly": 19091,
      "popular generative": 25117,
      "overall observe": 23909,
      "behaviors models": 3328,
      "users requirements": 34702,
      "existing zeroshot": 10329,
      "tasks little": 32406,
      "little research": 18052,
      "translation particularly": 33843,
      "data limited": 6757,
      "given document": 12747,
      "report empirically": 28114,
      "use various": 34580,
      "various prompts": 35136,
      "different paradigms": 8116,
      "endtoend pipeline": 9439,
      "compared finetuned": 5133,
      "using multidimensional": 34846,
      "interact users": 15381,
      "task finetuned": 32124,
      "process based": 26051,
      "performance datasets": 24562,
      "aspects including": 2575,
      "approach efficiently": 2267,
      "transferable prompt": 33687,
      "efficiently adapt": 9069,
      "cases despite": 4199,
      "despite tuning": 7824,
      "transfer capability": 33670,
      "tasks adaptation": 32233,
      "data does": 6684,
      "visual inputs": 35336,
      "information visual": 14926,
      "input experiments": 15010,
      "models extended": 21244,
      "range social": 27211,
      "use human": 34536,
      "approach uses": 2355,
      "model reward": 20762,
      "various kinds": 35102,
      "control format": 6049,
      "architectures focus": 2462,
      "help bridge": 13504,
      "gap different": 12088,
      "signals different": 29944,
      "effectiveness efficiency": 8943,
      "tasks dynamic": 32303,
      "reinforcement learningbased": 27847,
      "determine optimal": 7898,
      "time request": 33139,
      "synthetic realworld": 31859,
      "alternative approaches": 1850,
      "specifically demonstrate": 30733,
      "use small": 34570,
      "approach help": 2293,
      "output models": 23876,
      "compared different": 5130,
      "prompting multilingual": 26391,
      "recent proliferation": 27545,
      "systems generating": 31899,
      "capabilities generating": 4021,
      "chatgpt generates": 4473,
      "generates fluent": 12401,
      "enable users": 9289,
      "user experiences": 34648,
      "data user": 6904,
      "works phases": 35819,
      "phases phase": 24920,
      "taskspecific training": 32570,
      "final phase": 11180,
      "experimental analysis": 10383,
      "observed medical": 23240,
      "addition model": 1242,
      "allowing model": 1806,
      "led astray": 17683,
      "features significantly": 11038,
      "standard datasets": 30873,
      "models core": 21112,
      "compared gpt3": 5136,
      "gpt3 despite": 12990,
      "times smaller": 33165,
      "dataset conducted": 6958,
      "efficiently finetune": 9072,
      "instructionfollowing model": 15235,
      "hour finetuning": 13765,
      "tokens higher": 33231,
      "preserves pretrained": 25607,
      "extended multimodal": 10661,
      "multimodal instructions": 22348,
      "benchmarks furthermore": 3443,
      "models vit": 22119,
      "tasks demonstrating": 32290,
      "task acquiring": 32073,
      "second step": 29328,
      "key metrics": 15778,
      "study demonstrate": 31314,
      "training mechanism": 33560,
      "follow uniform": 11679,
      "quality using": 26983,
      "modelgenerated responses": 20884,
      "quality responses": 26976,
      "quality degradation": 26952,
      "documents models": 8519,
      "generate hallucinated": 12279,
      "approximation fisher": 2419,
      "fisher information": 11569,
      "information matrix": 14887,
      "multiple datasets": 22385,
      "method extended": 19918,
      "success heavily": 31511,
      "potential building": 25247,
      "challenges achieving": 4333,
      "maintaining consistency": 19423,
      "consistency human": 5731,
      "comprehensive studies": 5391,
      "multiagent systems": 22268,
      "human instructions": 13826,
      "resulting suboptimal": 28563,
      "chatgpt shown": 4496,
      "pipeline automatically": 24964,
      "model resulting": 20756,
      "furthermore propose": 12010,
      "new technique": 22858,
      "feedback improve": 11063,
      "document generation": 8503,
      "generate different": 12272,
      "applying pretrained": 2221,
      "consisting 100": 5760,
      "gpt4 outperformed": 13106,
      "observed human": 23238,
      "groups evaluating": 13305,
      "demonstrated surprising": 7554,
      "outperform chatgpt": 23767,
      "arithmetic tasks": 2505,
      "llama various": 18150,
      "codes released": 4856,
      "multimodal neural": 22365,
      "diversity data": 8477,
      "data difficulty": 6677,
      "captioning datasets": 4137,
      "images videos": 14091,
      "used variety": 34635,
      "poses challenge": 25166,
      "challenge diverse": 4310,
      "issues present": 15671,
      "datasets approach": 7066,
      "input position": 15020,
      "embeddings pretrained": 9145,
      "approach offers": 2320,
      "llms t5": 18988,
      "embeddings language": 9142,
      "generation technique": 12618,
      "studies multilingual": 31277,
      "strong supervised": 31193,
      "better task": 3629,
      "model demonstrated": 20456,
      "model accessible": 20341,
      "models efficiently": 21185,
      "margin work": 19586,
      "users preference": 34697,
      "works used": 35829,
      "taken account": 32023,
      "tasks prompt": 32459,
      "information news": 14892,
      "different users": 8156,
      "vast number": 35189,
      "diverse corpora": 8419,
      "specifically pretrain": 30752,
      "using original": 34868,
      "original pretraining": 23718,
      "systems struggle": 31919,
      "small mediumsized": 30356,
      "datasets results": 7170,
      "tasks leads": 32396,
      "investigated models": 15602,
      "visual instruction": 35337,
      "tuning instruction": 33983,
      "using machinegenerated": 34833,
      "capabilities new": 4052,
      "present attempt": 25512,
      "attempt use": 2705,
      "data introduce": 6742,
      "large multimodal": 17245,
      "multimodal model": 22362,
      "llm generalpurpose": 18312,
      "early experiments": 8776,
      "multimodal instructionfollowing": 22347,
      "llava gpt4": 18244,
      "gpt4 generated": 13081,
      "revolutionizing natural": 28854,
      "exponentially increasing": 10629,
      "unidirectional attention": 34320,
      "techniques employed": 32636,
      "analysis furthermore": 1923,
      "encoder models": 9353,
      "autoregressive causal": 2936,
      "heads task": 13482,
      "trained joint": 33403,
      "graph information": 13221,
      "abilities directly": 624,
      "visionlanguage models": 35315,
      "technical details": 32605,
      "multimodal generation": 22344,
      "encoder frozen": 9348,
      "vicuna using": 35256,
      "projection layer": 26250,
      "work time": 35794,
      "detailed image": 7838,
      "collected dataset": 4922,
      "image semantic": 14075,
      "models fms": 21287,
      "contexts minimal": 5944,
      "tuning code": 33969,
      "llm backbone": 18274,
      "exploratory research": 10568,
      "presented major": 25566,
      "token position": 33196,
      "transformers language": 33786,
      "shown stateoftheart": 29917,
      "known suffer": 15937,
      "specific issue": 30697,
      "indepth evaluation": 14677,
      "decoders gpt2": 7267,
      "propose methods": 26530,
      "methods random": 20083,
      "process results": 26084,
      "results improvement": 28631,
      "improvement approx": 14328,
      "llms showcased": 18936,
      "models frequently": 21297,
      "powerful opensource": 25349,
      "finetuning alignment": 11370,
      "alignment domainspecific": 1756,
      "domainspecific instructions": 8650,
      "comprehensive dataset": 5363,
      "recently popular": 27613,
      "training llm": 33553,
      "demonstrates potential": 7563,
      "llms generalize": 18632,
      "fusion strategy": 12026,
      "visual tokens": 35359,
      "llm layers": 18331,
      "imagetext pairs": 14094,
      "achieves strong": 1074,
      "dataset inference": 7003,
      "incorporate additional": 14569,
      "expert models": 10512,
      "image understanding": 14077,
      "costs compared": 6268,
      "software engineering": 30455,
      "algorithms generate": 1718,
      "powerful capabilities": 25336,
      "fields including": 11155,
      "applications models": 2169,
      "models discuss": 21160,
      "discuss limitations": 8294,
      "point future": 25064,
      "tuning framework": 33979,
      "align large": 1728,
      "potential use": 25304,
      "recommendation systems": 27651,
      "data end": 6690,
      "framework aligning": 11827,
      "fewer 100": 11085,
      "100 samples": 38,
      "furthermore finetuned": 12002,
      "llm exhibits": 18299,
      "final results": 11184,
      "llms propose": 18875,
      "efficient search": 9056,
      "approach surpasses": 2343,
      "results llama2": 28640,
      "demonstrate efficiency": 7452,
      "robustness code": 28943,
      "visual prompt": 35347,
      "developing new": 7946,
      "multimodal llm": 22360,
      "existing llm": 10289,
      "alternative solution": 1856,
      "existing mllms": 10297,
      "solution reduce": 30478,
      "transfer different": 33672,
      "different llm": 8094,
      "significantly speed": 30087,
      "finally showcase": 11204,
      "mllms including": 20298,
      "released llama": 27924,
      "llms automatically": 18440,
      "generations different": 12642,
      "smallscale study": 30406,
      "study compare": 31306,
      "medical domains": 19768,
      "undesirable biases": 34292,
      "second use": 29329,
      "responses users": 28515,
      "desirable responses": 7759,
      "develop ai": 7909,
      "lines human": 18002,
      "including textdavinci003": 14521,
      "using vicuna": 34939,
      "vicuna large": 35250,
      "fewshot capabilities": 11098,
      "problems paper": 26030,
      "performance shot": 24751,
      "vicuna multiple": 35255,
      "generally requires": 12244,
      "requires large": 28256,
      "range nlp": 27205,
      "scientific domain": 29250,
      "recent advancement": 27488,
      "recognition tasks": 27641,
      "plms based": 25040,
      "works limited": 35815,
      "demonstrated significant": 7548,
      "introduce instruction": 15511,
      "required training": 28233,
      "huggingface transformers": 13788,
      "models customized": 21120,
      "customized training": 6565,
      "inference pipelines": 14798,
      "representations query": 28173,
      "available paper": 2995,
      "languages training": 16917,
      "training effective": 33506,
      "method results": 19971,
      "results experiments": 28610,
      "gpt4 based": 13059,
      "unfortunately model": 34314,
      "multimodal information": 22345,
      "dataset conduct": 6957,
      "multiround dialogue": 22437,
      "generating detailed": 12419,
      "lowrank adapter": 19303,
      "tuning make": 33997,
      "enhance ability": 9501,
      "effectively improves": 8921,
      "biomedical text": 3748,
      "based encoder": 3154,
      "legal domain": 17696,
      "complex datasets": 5272,
      "attention research": 2740,
      "learn underlying": 17515,
      "different approach": 8047,
      "llms key": 18732,
      "llms understand": 19017,
      "instead using": 15126,
      "general instruction": 12167,
      "manually design": 19569,
      "learn language": 17510,
      "plms existing": 25045,
      "image encoder": 14065,
      "encoder visionlanguage": 9358,
      "models vlms": 22121,
      "pretrained vlms": 25779,
      "large visionlanguage": 17291,
      "models enhanced": 21206,
      "exceptional proficiency": 10174,
      "diverse linguistic": 8437,
      "minigpt4 llava": 20177,
      "established benchmarks": 9774,
      "learning reason": 17641,
      "benchmark compare": 3358,
      "generalizability llmbased": 12201,
      "promising potential": 26294,
      "struggle produce": 31243,
      "parameters gptneo": 24254,
      "hundreds millions": 13945,
      "millions parameters": 20168,
      "attention work": 2745,
      "generated gpt35": 12361,
      "models suggest": 22028,
      "models output": 21785,
      "score model": 29273,
      "language capabilities": 16046,
      "knowledge enhancement": 15844,
      "task automatically": 32083,
      "approach grounded": 2291,
      "knowledge different": 15829,
      "used generating": 34604,
      "llms exploit": 18591,
      "gpt3 llama": 13000,
      "induce implicit": 14731,
      "code open": 4794,
      "case semantic": 4194,
      "propose using": 26584,
      "given collection": 12742,
      "collection usage": 4934,
      "usage examples": 34503,
      "demonstrate resulting": 7489,
      "finetuning prompting": 11498,
      "open pretrained": 23411,
      "transformers opt": 33792,
      "entails finetuning": 9617,
      "outofdomain tasks": 23753,
      "techniques comprehensive": 32631,
      "significant impact": 29987,
      "explicitly tailored": 10550,
      "bias potential": 3659,
      "gap paper": 12101,
      "distillation methods": 8345,
      "results imply": 28627,
      "misinformation detection": 20213,
      "spreading misinformation": 30830,
      "domains remains": 8636,
      "task misinformation": 32156,
      "good data": 12819,
      "initial model": 14960,
      "performance domain": 24573,
      "linguistic characteristics": 18008,
      "languages introduce": 16881,
      "evaluation conducted": 9931,
      "families demonstrate": 10966,
      "chatgpt openai": 4490,
      "googles bard": 12832,
      "products services": 26173,
      "data experiments": 6705,
      "demands large": 7416,
      "expensive propose": 10366,
      "strategy leverages": 31126,
      "model datasets": 20452,
      "techniques perform": 32656,
      "leads competitive": 17490,
      "performance quality": 24728,
      "architectures using": 2474,
      "using roberta": 34904,
      "morphological syntactic": 22220,
      "analysis existing": 1920,
      "novel generative": 23083,
      "resources large": 28438,
      "existing efforts": 10270,
      "relied supervised": 27964,
      "generalization new": 12223,
      "llms capability": 18457,
      "attempt investigate": 2704,
      "investigate feasibility": 15583,
      "extensive ablation": 10675,
      "distillation proprietary": 8346,
      "chatgpt work": 4499,
      "prompts use": 26445,
      "model aligned": 20362,
      "aligned language": 1739,
      "outperforms recent": 23849,
      "opensourced models": 23556,
      "framework code": 11835,
      "parameters present": 24277,
      "present challenges": 25518,
      "deployment previous": 7658,
      "studies try": 31288,
      "cot finetuning": 6280,
      "data contains": 6658,
      "capabilities work": 4084,
      "reasoning program": 27440,
      "baselines significantly": 3270,
      "smaller scale": 30395,
      "dataset evaluating": 6984,
      "exciting recent": 10185,
      "power robustness": 25329,
      "predict future": 25408,
      "contrast large": 6013,
      "flan collection": 11587,
      "finetuning flant5": 11405,
      "lms better": 19074,
      "terms zeroshot": 32753,
      "task accuracy": 32071,
      "furthermore instruction": 12003,
      "language modelslms": 16762,
      "concepts essential": 5531,
      "design benchmark": 7698,
      "objects ii": 23215,
      "does apply": 8523,
      "clip blip": 4663,
      "visual representation": 35355,
      "valuable source": 35018,
      "inspired propose": 15098,
      "propose distillation": 26505,
      "detecting factual": 7855,
      "including vanilla": 14527,
      "llms able": 18407,
      "learning incontext": 17586,
      "llms providing": 18881,
      "underlying mechanism": 34158,
      "compression technique": 5428,
      "llama outperforms": 18138,
      "generated dataset": 12350,
      "nearperfect accuracy": 22604,
      "evaluation effectiveness": 9943,
      "hallucinate wrong": 13368,
      "12 billion": 88,
      "improve factuality": 14267,
      "used finetune": 34601,
      "alpaca experimental": 1828,
      "highquality instructional": 13693,
      "finetuning instruction": 11422,
      "designed diverse": 7726,
      "largescale dataset": 17346,
      "leading opensource": 17480,
      "remarkable advancements": 28028,
      "selecting appropriate": 29385,
      "objective investigate": 23206,
      "compare various": 5115,
      "performs poorly": 24852,
      "poorly context": 25107,
      "highquality examples": 13688,
      "examples exhibit": 10122,
      "test examples": 32767,
      "evaluation finegrained": 9947,
      "quality language": 26971,
      "high correlation": 13560,
      "including based": 14460,
      "gpt3 gpt4": 12998,
      "loss quality": 19252,
      "latency cost": 17401,
      "gpt4 compared": 13067,
      "robustness finetuned": 28946,
      "tasks shown": 32500,
      "shown effective": 29872,
      "work studying": 35791,
      "bert finetuned": 3503,
      "layers using": 17449,
      "study robustness": 31391,
      "text perturbations": 32916,
      "like summarization": 17894,
      "generation question": 12585,
      "types input": 34064,
      "perturbation models": 24909,
      "aim generate": 1643,
      "texts lack": 32995,
      "lack clarity": 15976,
      "based specific": 3224,
      "baselines based": 3262,
      "observed finetuned": 23236,
      "models address": 20961,
      "propose explicit": 26510,
      "gains transformer": 12074,
      "decoderonly large": 7257,
      "queries learning": 27023,
      "reverse engineering": 28823,
      "development design": 7953,
      "developed predict": 7932,
      "generated code": 12347,
      "designed establish": 7730,
      "evaluations stateoftheart": 10040,
      "different transfer": 8153,
      "reveal significant": 28806,
      "significant room": 30024,
      "examples analysis": 10117,
      "avenues future": 3006,
      "candidates based": 3995,
      "significant time": 30027,
      "time effort": 33120,
      "tulu llama2": 33954,
      "variable values": 35035,
      "involves exploring": 15628,
      "model reasoning": 20743,
      "llm agent": 18265,
      "plan generation": 24994,
      "typically covered": 34077,
      "problem models": 25996,
      "models access": 20939,
      "use proposed": 34563,
      "closed models": 4678,
      "models struggling": 22018,
      "intelligence existing": 15355,
      "image language": 14071,
      "algorithm help": 1708,
      "single multimodal": 30214,
      "llm called": 18278,
      "demonstrate competitive": 7441,
      "training hours": 33528,
      "extremescale language": 10805,
      "tasks toxicity": 32532,
      "brings significant": 3880,
      "models outperforms": 21784,
      "highlight potential": 13635,
      "comprehensive model": 5387,
      "leveraging advanced": 17778,
      "data enhance": 6693,
      "documentation essential": 8511,
      "efficiency text": 9015,
      "parallel context": 24166,
      "context windows": 5929,
      "limitations evaluation": 17918,
      "positional embedding": 25187,
      "inference based": 14762,
      "enabling language": 9323,
      "harnessing power": 13467,
      "literature paper": 18043,
      "gpt35 achieve": 13015,
      "performance gpt4": 24620,
      "levels complexity": 17740,
      "validity generated": 35009,
      "abilities various": 646,
      "integrates llms": 15328,
      "specifically design": 30734,
      "generation reasoning": 12589,
      "editing various": 8830,
      "study task": 31401,
      "experiment different": 10377,
      "participants evaluate": 24327,
      "different modalities": 8103,
      "text video": 32965,
      "models word": 22132,
      "current paper": 6522,
      "media data": 19759,
      "common words": 5020,
      "impact model": 14131,
      "tasks researchers": 32486,
      "approach captures": 2247,
      "original input": 23709,
      "mechanism evaluate": 19750,
      "backpack language": 3065,
      "neural architecture": 22720,
      "modeling performance": 20904,
      "sense vectors": 29509,
      "linear combination": 17985,
      "linear projection": 17993,
      "change models": 4407,
      "finally present": 11200,
      "gender bias": 12149,
      "process work": 26088,
      "model synthesize": 20819,
      "fusion layer": 12025,
      "analyzing generated": 2003,
      "generated output": 12376,
      "model reveal": 20761,
      "primary challenge": 25919,
      "lack understanding": 16009,
      "propose explore": 26511,
      "furthermore investigate": 12005,
      "makes mistakes": 19492,
      "challenging work": 4404,
      "building better": 3919,
      "improves generalization": 14378,
      "trained small": 33426,
      "emotional state": 9205,
      "joint prediction": 15714,
      "result attain": 28540,
      "inference times": 14818,
      "new era": 22798,
      "results llms": 28641,
      "making model": 19511,
      "determine final": 7897,
      "approach successfully": 2342,
      "successfully mitigates": 31543,
      "bias resulting": 3662,
      "augmenting large": 2825,
      "languages making": 16896,
      "training llama": 33551,
      "largescale parallel": 17371,
      "chatgpt google": 4474,
      "tools allow": 33268,
      "based preceding": 3207,
      "addresses challenges": 1363,
      "benchmark furthermore": 3388,
      "demonstrates robustness": 7565,
      "prompt efficiency": 26320,
      "reducing model": 27756,
      "gpt35 7b": 13014,
      "poses challenges": 25167,
      "assess existing": 2592,
      "manual automatic": 19555,
      "improve explainability": 14265,
      "concerns data": 5541,
      "llms augmented": 18438,
      "propose neural": 26535,
      "performance small": 24755,
      "larger parameters": 17335,
      "potential tool": 25302,
      "sophisticated prompt": 30527,
      "typically rely": 34080,
      "prohibitive computational": 26236,
      "data address": 6594,
      "various multimodal": 35123,
      "multimodal contexts": 22339,
      "unseen tools": 34445,
      "alpaca 7b": 1822,
      "developed model": 7930,
      "participants able": 24326,
      "critical research": 6392,
      "research problem": 28347,
      "lms typically": 19118,
      "gpt llama2": 12855,
      "features used": 11043,
      "key innovation": 15773,
      "decisionmaking process": 7237,
      "informative features": 14932,
      "furthermore method": 12006,
      "times improvement": 33163,
      "method extends": 19919,
      "excessive memory": 10179,
      "memory overhead": 19823,
      "overhead paper": 23937,
      "focus exploring": 11647,
      "based observations": 3204,
      "tasks short": 32498,
      "tasks apply": 32245,
      "instructiontuning language": 15305,
      "observation propose": 23219,
      "like alpaca": 17844,
      "rapid progress": 27253,
      "costefficient approach": 6261,
      "use gpt4": 34534,
      "finetune large": 11285,
      "method specifically": 19976,
      "openended instruction": 23458,
      "multimodal research": 22366,
      "api calls": 2102,
      "contexts given": 5939,
      "simple sequences": 30163,
      "models unable": 22089,
      "investigates effectiveness": 15605,
      "training scheme": 33605,
      "datasets considerable": 7082,
      "leverages recent": 17773,
      "tasks growing": 32352,
      "effectiveness flexibility": 8946,
      "understand adaptability": 34187,
      "claude vicuna": 4637,
      "introduce additional": 15497,
      "requiring finetuning": 28272,
      "algorithm significantly": 1714,
      "researchers developers": 28374,
      "develop advanced": 7908,
      "multilingual nature": 22322,
      "capability plms": 4100,
      "using parameterefficient": 34869,
      "verification tasks": 35210,
      "experiments public": 10471,
      "framework empowers": 11844,
      "capability understanding": 4104,
      "audio encoders": 2777,
      "audio signals": 2780,
      "pretrained audio": 25626,
      "audio encoder": 2776,
      "content generate": 5859,
      "auditory information": 2788,
      "llms difficult": 18541,
      "inferencetime approach": 14827,
      "semantic constraints": 29452,
      "syntactic constraints": 31814,
      "domainadaptive pretraining": 8607,
      "adaptation methods": 1185,
      "introduce inferencetime": 15509,
      "truthfulness large": 33940,
      "model activations": 20354,
      "number attention": 23139,
      "technique data": 32619,
      "data efficient": 6689,
      "reasoning traces": 27462,
      "use state": 34572,
      "constrain generation": 5778,
      "critical training": 6397,
      "million chinese": 20163,
      "decoderonly model": 7262,
      "emphasizes need": 9213,
      "tools help": 33270,
      "sentences abstracts": 29550,
      "method finding": 19920,
      "task specifically": 32194,
      "dataset does": 6978,
      "emphasizing importance": 9216,
      "task code": 32093,
      "specific llm": 30703,
      "user query": 34668,
      "prompts generated": 26420,
      "trained llama": 33407,
      "demonstrated ability": 7515,
      "learning requires": 17644,
      "scarcity data": 29189,
      "settings findings": 29734,
      "setting instruction": 29723,
      "25 downstream": 271,
      "conduct analysis": 5584,
      "learning additionally": 17532,
      "additionally observe": 1294,
      "contrary previous": 6009,
      "generation coherent": 12473,
      "like large": 17879,
      "align human": 1726,
      "components model": 5314,
      "achieving optimal": 1098,
      "question using": 27077,
      "distributional properties": 8400,
      "time hypothesis": 33127,
      "text sampling": 32937,
      "text findings": 32858,
      "test perplexity": 32779,
      "domainspecific settings": 8656,
      "contrast traditional": 6018,
      "consistently underperforms": 5758,
      "reveal llm": 28802,
      "taskagnostic manner": 32213,
      "performance outperforms": 24707,
      "augment pretrained": 2792,
      "specifically identify": 30745,
      "identify address": 14003,
      "efficiency costeffectiveness": 8999,
      "systems conduct": 31890,
      "designs existing": 7756,
      "models latest": 21436,
      "simple linear": 30156,
      "process approach": 26050,
      "domain large": 8571,
      "vanilla finetuning": 35029,
      "analysis results": 1956,
      "llms step": 18969,
      "including opensource": 14510,
      "tuning approach": 33967,
      "development healthcare": 7955,
      "used models": 34614,
      "challenges paper": 4365,
      "research investigating": 28329,
      "domains computer": 8615,
      "hierarchical data": 13542,
      "results similar": 28684,
      "impact performance": 14136,
      "using parameters": 34872,
      "great improvement": 13251,
      "necessary achieve": 22607,
      "results solving": 28688,
      "efficient communication": 9029,
      "human understanding": 13872,
      "tasks focus": 32341,
      "compelling evidence": 5208,
      "substantial improvement": 31466,
      "direction future": 8224,
      "information transformerbased": 14922,
      "amr parsing": 1899,
      "abstract meaning": 770,
      "representation amr": 28139,
      "text current": 32840,
      "approaches based": 2368,
      "sentence paper": 29540,
      "method explores": 19917,
      "data release": 6831,
      "finetuning final": 11404,
      "fair comparison": 10923,
      "determine practical": 7899,
      "model extensive": 20506,
      "significant difference": 29977,
      "large vision": 17287,
      "experiments largescale": 10457,
      "llama code": 18087,
      "llm llm": 18336,
      "llm used": 18377,
      "domain prompt": 8585,
      "prompt methods": 26335,
      "methods effectively": 20023,
      "instructions like": 15261,
      "tackle propose": 32001,
      "highly accurate": 13657,
      "detection performance": 7876,
      "language using": 16852,
      "prediction approach": 25421,
      "surpassing counterparts": 31753,
      "remains relatively": 28010,
      "relatively modest": 27890,
      "superior capability": 31646,
      "furthermore recent": 12011,
      "use publicly": 34564,
      "dataset prompt": 7024,
      "achieving accuracy": 1083,
      "model natural": 20652,
      "online content": 23363,
      "outperform finetuned": 23774,
      "benchmarks using": 3476,
      "best approach": 3554,
      "using prompt": 34884,
      "possible achieve": 25212,
      "domain recent": 8587,
      "advancements language": 1464,
      "led emergence": 17685,
      "capabilities wide": 4082,
      "models fewer": 21266,
      "aims provide": 1671,
      "based extensive": 3157,
      "ner tasks": 22681,
      "llms fewshot": 18605,
      "findings introduce": 11240,
      "relevant examples": 27940,
      "model generative": 20550,
      "unexplored paper": 34300,
      "innovative approach": 14995,
      "novel llm": 23094,
      "llm directly": 18291,
      "ability interpret": 686,
      "human intent": 13828,
      "llms align": 18429,
      "machine generated": 19341,
      "despite huge": 7785,
      "checkpoint model": 4519,
      "perception models": 24463,
      "number instructions": 23147,
      "generation prowess": 12582,
      "modular framework": 22186,
      "tasks efficiently": 32305,
      "methods exhibit": 20030,
      "method does": 19903,
      "input video": 15037,
      "model joint": 20596,
      "compared supervised": 5175,
      "supervised methods": 31687,
      "methods evaluated": 20028,
      "potential aligning": 25238,
      "requirements associated": 28236,
      "solution selectively": 30479,
      "finetuning small": 11528,
      "adaptation study": 1193,
      "adapter layer": 1200,
      "additionally propose": 1297,
      "datasets comparing": 7079,
      "tasks observe": 32434,
      "abilities specific": 642,
      "knowledge especially": 15846,
      "advance field": 1418,
      "integration large": 15340,
      "enabling better": 9317,
      "textbased large": 32973,
      "training smaller": 33618,
      "data conduct": 6652,
      "model multimodal": 20648,
      "largescale transformerbased": 17385,
      "architectural changes": 2434,
      "emergent cognitive": 9185,
      "enhance problemsolving": 9526,
      "based task": 3229,
      "compared using": 5180,
      "effectively reduces": 8929,
      "capabilities additionally": 4000,
      "comparative experiments": 5099,
      "development code": 7951,
      "remarkable ability": 28026,
      "limiting applicability": 17976,
      "method automatically": 19883,
      "automatically generates": 2915,
      "diverse multilingual": 8440,
      "finetuning assess": 11374,
      "collect existing": 4915,
      "surpasses opensource": 31747,
      "corpus pretraining": 6186,
      "smaller training": 30398,
      "based statistical": 3225,
      "various prompt": 35134,
      "lms llms": 19097,
      "results strategy": 28690,
      "model outperform": 20668,
      "prior approaches": 25935,
      "approaches utilize": 2403,
      "established baselines": 9773,
      "ability various": 730,
      "language especially": 16069,
      "incorporating multimodal": 14577,
      "including image": 14494,
      "image video": 14078,
      "effectiveness generating": 8948,
      "ability ground": 681,
      "user experience": 34647,
      "audio language": 2779,
      "twostage training": 34047,
      "open foundation": 23393,
      "release llama": 27909,
      "llama 2chat": 18063,
      "cases models": 4205,
      "helpfulness safety": 13520,
      "order enable": 23671,
      "work contribute": 35682,
      "responsible development": 28521,
      "help identify": 13505,
      "potentially support": 25315,
      "discovery paper": 8275,
      "engine generate": 9459,
      "models static": 22011,
      "sensitive perturbations": 29517,
      "incorrect irrelevant": 14585,
      "metrics evaluating": 20136,
      "overall work": 23917,
      "demonstrates training": 7578,
      "identifying interpretable": 14021,
      "explanation using": 10539,
      "like clip": 17857,
      "models 20": 20926,
      "tasks finally": 32330,
      "33 billion": 330,
      "parameters small": 24291,
      "a100 80gb": 615,
      "exhibiting impressive": 10241,
      "distinct advantage": 8363,
      "vicuna llama": 35252,
      "using vanilla": 34935,
      "able solve": 754,
      "prepending sequence": 25501,
      "studies investigate": 31273,
      "maintain original": 19416,
      "embeddings results": 9147,
      "biomedical natural": 3745,
      "certain tasks": 4281,
      "variety applications": 35057,
      "llms field": 18606,
      "evaluation multiple": 9978,
      "alpaca alpacalora": 1823,
      "flant5 gpt35": 11594,
      "finetuning results": 11512,
      "experiments instruction": 10451,
      "finetuning significantly": 11524,
      "tasks simultaneously": 32506,
      "outperform best": 23766,
      "balanced accuracy": 3083,
      "model conduct": 20437,
      "study llms": 31359,
      "tasks illustrating": 32357,
      "illustrating promising": 14051,
      "racial gender": 27145,
      "various modalities": 35118,
      "generative visionlanguage": 12713,
      "applications data": 2147,
      "continue pretraining": 5987,
      "evaluate datasets": 9824,
      "including novel": 14507,
      "problems furthermore": 26027,
      "research paper": 28339,
      "hallucinations large": 13391,
      "reduce hallucinations": 27714,
      "leading llms": 17478,
      "including text": 14520,
      "promoting transparency": 26304,
      "development safer": 7973,
      "safer reliable": 29040,
      "exploring capabilities": 10615,
      "text space": 32946,
      "use autoregressive": 34514,
      "recipe training": 27631,
      "based multimodal": 3200,
      "results recently": 28668,
      "shot setting": 29832,
      "art natural": 2516,
      "open resources": 23420,
      "specifically tuned": 30759,
      "users prompts": 34700,
      "instruction prompts": 15175,
      "ability handle": 682,
      "detect using": 7852,
      "highly systematic": 13670,
      "training evaluating": 33512,
      "evaluations finetuned": 10029,
      "openais gpt3": 23444,
      "gpt3 llms": 13001,
      "framework training": 11898,
      "large autoregressive": 16929,
      "local training": 19136,
      "local llms": 19132,
      "reduced precision": 27730,
      "datasets demonstrating": 7094,
      "presents effective": 25580,
      "data influence": 6736,
      "parameter count": 24176,
      "improving model": 14415,
      "model performances": 20707,
      "rejection sampling": 27851,
      "samples multiple": 29085,
      "llm foundation": 18309,
      "models emergent": 21189,
      "sentences task": 29560,
      "hallucinations llms": 13394,
      "provide results": 26727,
      "generation test": 12620,
      "improvement using": 14350,
      "study aim": 31296,
      "construction model": 5811,
      "construct largescale": 5801,
      "text descriptions": 32844,
      "comprehensively assess": 5397,
      "multimodal foundation": 22342,
      "surpassing existing": 31754,
      "distillation large": 8340,
      "models alpaca": 20978,
      "original llms": 23712,
      "applications paper": 2170,
      "extraction using": 10776,
      "distilled smaller": 8355,
      "benchmark date": 3376,
      "attains remarkable": 2699,
      "tens thousands": 32719,
      "models facilitate": 21254,
      "general use": 12189,
      "llms rely": 18902,
      "training pipeline": 33587,
      "given unique": 12781,
      "baselines various": 3274,
      "safety code": 29045,
      "capabilities increasingly": 4026,
      "collect dataset": 4914,
      "commercial opensource": 4993,
      "including variations": 14528,
      "better alignment": 3589,
      "advance language": 1419,
      "role current": 28955,
      "datasets extensive": 7114,
      "remains understudied": 28020,
      "zeroshot abilities": 35950,
      "event detection": 10046,
      "method gpt2": 19928,
      "limited size": 17966,
      "detection incontext": 7867,
      "finetuning note": 11464,
      "examples model": 10136,
      "prompting incontext": 26377,
      "substantial parameter": 31471,
      "enhanced reasoning": 9538,
      "method designed": 19899,
      "models play": 21817,
      "potential models": 25279,
      "content specific": 5873,
      "study investigated": 31347,
      "gpt4 achieved": 13052,
      "implications future": 14174,
      "llms introduces": 18729,
      "offering unified": 23301,
      "unified solution": 34338,
      "solution address": 30468,
      "privacy data": 25952,
      "instructionfollowing language": 15229,
      "models comprehensively": 21079,
      "comprehensively understand": 5399,
      "performance largely": 24649,
      "interpreting complex": 15468,
      "inherent knowledge": 14949,
      "additionally compare": 1274,
      "used conjunction": 34589,
      "general domainspecific": 12163,
      "object detection": 23199,
      "models openais": 21769,
      "ability detect": 660,
      "approach adaptively": 2230,
      "lowrank structure": 19308,
      "surpassing stateoftheart": 31760,
      "reasoning boost": 27381,
      "tasks chainofthought": 32259,
      "learning multimodal": 17615,
      "hypothesize large": 13967,
      "automated detection": 2859,
      "types models": 34068,
      "bert pretrained": 3523,
      "llama closedsource": 18086,
      "versions gpt3": 35235,
      "recently developed": 27591,
      "including model": 14504,
      "accuracy privacy": 903,
      "layer transformer": 17432,
      "insights human": 15077,
      "identify social": 14015,
      "models extract": 21251,
      "extremely valuable": 10803,
      "bestperforming finetuned": 3583,
      "models likely": 21461,
      "settings models": 29739,
      "models enhance": 21205,
      "ways using": 35453,
      "showing potential": 29861,
      "task ablation": 32070,
      "exploring instruction": 10617,
      "models systematic": 22038,
      "review automation": 28829,
      "traditional approaches": 33342,
      "process explore": 26059,
      "compression inference": 5416,
      "noteworthy compression": 23038,
      "achieves compression": 1044,
      "models demonstrating": 21139,
      "allows direct": 1812,
      "direct application": 8211,
      "models quantitative": 21881,
      "zero oneshot": 35944,
      "model prediction": 20715,
      "using gptj": 34793,
      "points code": 25071,
      "conduct set": 5619,
      "experiments explore": 10441,
      "google cloud": 12827,
      "human natural": 13847,
      "gap language": 12094,
      "users easily": 34686,
      "modalities natural": 20324,
      "models reinforced": 21916,
      "outperforms chatgpt35": 23813,
      "chatgpt35 claude": 4501,
      "public httpsgithubcomnlpxucanwizardlm": 26838,
      "extend large": 10653,
      "embeddings designed": 9139,
      "token count": 33188,
      "process empirical": 26057,
      "vqa benchmarks": 35395,
      "comprehensive multimodal": 5388,
      "comparing baseline": 5188,
      "llms employed": 18558,
      "demonstrate zeroshot": 7513,
      "chinese large": 4546,
      "rise popularity": 28888,
      "utilized training": 34965,
      "chatgpt conversations": 4462,
      "learning humans": 17582,
      "resulting limited": 28554,
      "demonstrates scalability": 7566,
      "era llms": 9702,
      "interactions mental": 15390,
      "customer service": 6558,
      "meticulously designed": 20117,
      "diverse scenarios": 8457,
      "pivotal step": 24981,
      "using information": 34799,
      "retrieval recommend": 28752,
      "engineering tasks": 9470,
      "potentially lead": 25313,
      "ml systems": 20291,
      "comparative studies": 5100,
      "significantly advanced": 30031,
      "tasks enhancing": 32313,
      "thoroughly investigated": 33077,
      "llama chatglm": 18083,
      "investigate effectiveness": 15581,
      "llms displayed": 18543,
      "evaluate quality": 9862,
      "generated different": 12352,
      "llms enhancing": 18566,
      "driven recent": 8736,
      "progress multimodal": 26217,
      "scarcity highquality": 29191,
      "highquality instructiontuning": 13695,
      "data requires": 6838,
      "mllms generate": 20296,
      "worth noting": 35843,
      "tend produce": 32709,
      "issue paper": 15657,
      "hallucination leveraging": 13380,
      "effectively enhances": 8917,
      "models lvlms": 21694,
      "set new": 29698,
      "settings zeroshot": 29746,
      "benchmarks instructiontuned": 3451,
      "demonstrates superiority": 7577,
      "according given": 847,
      "task difficulties": 32111,
      "guide language": 13347,
      "method attains": 19881,
      "attains stateoftheart": 2700,
      "relatively smaller": 27894,
      "recent chatgpt": 27511,
      "role optimizing": 28959,
      "scale context": 29131,
      "lms address": 19069,
      "32k 2k": 326,
      "lengths gpt4": 17715,
      "paper evaluates": 24041,
      "annotated corpus": 2016,
      "user goals": 34651,
      "demonstrated capability": 7517,
      "despite strong": 7816,
      "datasets lack": 7136,
      "anomaly detection": 2038,
      "lvlm generate": 19333,
      "impressive fewshot": 14237,
      "performance accuracy": 24515,
      "answering reasoning": 2072,
      "learn pretraining": 17512,
      "contrastive training": 6025,
      "additional modalities": 1259,
      "directly using": 8246,
      "models multimodal": 21741,
      "spread misinformation": 30828,
      "potentially leading": 25314,
      "relevant evidence": 27939,
      "accurately evaluate": 934,
      "experiments widely": 10503,
      "sufficient context": 31561,
      "findings implications": 11237,
      "language audio": 16043,
      "applications code": 2144,
      "make generated": 19470,
      "improves quality": 14391,
      "responses training": 28514,
      "sampling method": 29093,
      "generation training": 12623,
      "bias problem": 3660,
      "tools paper": 33273,
      "text achieves": 32814,
      "set code": 29676,
      "audio captioning": 2775,
      "attention propose": 2737,
      "text ii": 32893,
      "process involves": 26066,
      "text classifier": 32828,
      "classifier trained": 4626,
      "sentences present": 29558,
      "tuning present": 34006,
      "tuning method": 33998,
      "conditions including": 5579,
      "alignment training": 1782,
      "training training": 33638,
      "image features": 14066,
      "notably approach": 23029,
      "diverse modalities": 8438,
      "quality code": 26946,
      "mask token": 19607,
      "embeddings reduce": 9146,
      "tuning process": 34008,
      "parameter tuning": 24202,
      "parameter updating": 24203,
      "furthermore experiments": 12001,
      "using bertbased": 34741,
      "queries generated": 27022,
      "based information": 3177,
      "instructiontuned language": 15284,
      "level fkgl": 17732,
      "open closedsource": 23391,
      "provide empirical": 26696,
      "considered effective": 5717,
      "struggle perform": 31242,
      "approach pinpoint": 2326,
      "propose mechanism": 26526,
      "mechanism allows": 19749,
      "inference enabling": 14773,
      "attention layer": 2721,
      "decisions findings": 7241,
      "sensitive contextual": 29514,
      "strategies based": 31101,
      "underlying mechanics": 34157,
      "requiring complex": 28269,
      "empirically verify": 9252,
      "despite various": 7827,
      "mitigate forgetting": 20251,
      "light pressing": 17831,
      "pressing issue": 25616,
      "different layers": 8091,
      "layers transformer": 17447,
      "tradeoffs propose": 33339,
      "model layers": 20606,
      "process known": 26067,
      "architecture trained": 2454,
      "deployed models": 7635,
      "human expertise": 13816,
      "models realworld": 21896,
      "systems face": 31898,
      "issues related": 15674,
      "terms bleu": 32738,
      "significant obstacle": 29998,
      "enhances llms": 9546,
      "generate textual": 12333,
      "models attain": 20991,
      "improved truthfulness": 14320,
      "ethical alignment": 9803,
      "llama2chat 7b": 18218,
      "data releasing": 6834,
      "work explores": 35707,
      "autoregressive inference": 2941,
      "need extra": 22629,
      "model alignment": 20364,
      "training gradient": 33526,
      "gradient computation": 13186,
      "results evaluated": 28604,
      "gpt4 humans": 13085,
      "truthfulqa dataset": 33944,
      "data known": 6746,
      "llms prompt": 18874,
      "prompts proposed": 26436,
      "performance prompts": 24722,
      "gpt35 bard": 13017,
      "bard llama2": 3099,
      "provide novel": 26719,
      "increasingly crucial": 14635,
      "create challenging": 6347,
      "1024 tokens": 57,
      "objectives transformers": 23212,
      "applications introduce": 2157,
      "time maintaining": 33133,
      "applications reducing": 2175,
      "text spans": 32947,
      "models making": 21699,
      "models producing": 21857,
      "tackling challenge": 32006,
      "involves generating": 15630,
      "training robust": 33602,
      "represented training": 28192,
      "data expensive": 6701,
      "obtain paper": 23251,
      "textual corpora": 33023,
      "joint audio": 15711,
      "text existing": 32855,
      "existing new": 10300,
      "dataset key": 7007,
      "dataset specifically": 7040,
      "dataset comprises": 6955,
      "approach yielded": 2361,
      "yielded exceptional": 35915,
      "exceptional results": 10175,
      "tasks dataset": 32283,
      "domain need": 8579,
      "used construct": 34590,
      "audio events": 2778,
      "llava minigpt4": 18247,
      "performed using": 24832,
      "image resolution": 14074,
      "impact multimodal": 14135,
      "enhances model": 9547,
      "finetuning additionally": 11369,
      "additionally study": 1303,
      "highlights importance": 13652,
      "tuning improve": 33982,
      "makes stateoftheart": 19495,
      "forgetting multimodal": 11733,
      "following success": 11702,
      "llms vision": 19038,
      "problem multimodal": 25997,
      "forgetting mllms": 11732,
      "retain performance": 28717,
      "interestingly results": 15413,
      "text visual": 32966,
      "mllms demonstrate": 20295,
      "perform wide": 24509,
      "task explore": 32120,
      "conduct endtoend": 5600,
      "model built": 20404,
      "llm instruction": 18324,
      "takes input": 32033,
      "llm embeddings": 18296,
      "improvements approach": 14355,
      "llms billions": 18449,
      "report presents": 28123,
      "empirical observations": 9230,
      "observations inspire": 23222,
      "developers use": 7938,
      "checkpoints different": 4523,
      "training stages": 33622,
      "textual visual": 33041,
      "input semantic": 15024,
      "sft training": 29768,
      "experiments standard": 10483,
      "particular emphasis": 24337,
      "encoderonly decoderonly": 9375,
      "specifically present": 30751,
      "comprehensive assessment": 5355,
      "models emerging": 21191,
      "task especially": 32113,
      "approach llms": 2313,
      "finetuning stages": 11535,
      "achieve average": 943,
      "datasets performance": 7160,
      "novel training": 23121,
      "largescale realworld": 17380,
      "realworld llm": 27342,
      "containing million": 5835,
      "content including": 5862,
      "perform similarly": 24502,
      "safety benchmark": 29044,
      "success existing": 31509,
      "finetune llama2": 11289,
      "generate sentences": 12323,
      "combinatorial optimization": 4955,
      "optimization problem": 23634,
      "data structure": 6882,
      "sentences usually": 29562,
      "paper highlights": 24056,
      "generation especially": 12492,
      "method control": 19894,
      "task task": 32196,
      "use chatgpt": 34518,
      "length 512": 17704,
      "respectively additionally": 28455,
      "public health": 26837,
      "methods applied": 20006,
      "simplification biomedical": 30172,
      "detailed explanations": 7837,
      "predictions results": 25443,
      "results chatgpt": 28577,
      "analysis tasks": 1970,
      "fewshot prompts": 11120,
      "ensure reliability": 9606,
      "impressive capability": 14236,
      "including fully": 14480,
      "series llms": 29640,
      "series different": 29633,
      "datasets llms": 7145,
      "consistent considerable": 5736,
      "llms generalise": 18631,
      "set using": 29715,
      "proposed enable": 26597,
      "enable llms": 9288,
      "speech data": 30780,
      "area aims": 2477,
      "generation considering": 12477,
      "high fidelity": 13568,
      "process including": 26063,
      "rlhf large": 28908,
      "information context": 14857,
      "algorithm called": 1703,
      "vision instruction": 35301,
      "achieves remarkable": 1060,
      "94 performance": 600,
      "original image": 23708,
      "video use": 35266,
      "factually incorrect": 10897,
      "constraints specifically": 5792,
      "discover strong": 8268,
      "prompts study": 26442,
      "patterns predict": 24413,
      "mechanistic understanding": 19757,
      "involving multiple": 15635,
      "contributions propose": 6043,
      "llms openai": 18828,
      "failure modes": 10919,
      "implications application": 14173,
      "bases large": 3278,
      "need finetuning": 22630,
      "finetuning retraining": 11514,
      "query prompt": 27030,
      "mitigate challenges": 20249,
      "cover diverse": 6316,
      "diverse topics": 8469,
      "analysis comprising": 1914,
      "tasks previously": 32455,
      "previously thought": 25902,
      "comprehensive language": 5383,
      "alignment techniques": 1781,
      "utilizing code": 34969,
      "furthermore developed": 11994,
      "performance comparison": 24553,
      "translation engines": 33826,
      "generalpurpose large": 12250,
      "llms unified": 19019,
      "language bias": 16044,
      "bias llm": 3653,
      "expensive llm": 10360,
      "specialized models": 30674,
      "tuning llm": 33995,
      "compact model": 5066,
      "bias llms": 3654,
      "latency costs": 17402,
      "various evaluation": 35091,
      "failure cases": 10918,
      "works demonstrated": 35811,
      "general alignment": 12157,
      "domain specialization": 8593,
      "aligned models": 1741,
      "specialized domain": 30670,
      "unlabelled data": 34386,
      "data labeled": 6747,
      "reduce hallucination": 27713,
      "offers effective": 23307,
      "llm different": 18290,
      "combined form": 4959,
      "efficiency terms": 9014,
      "gained considerable": 12061,
      "considerable attention": 5705,
      "modeling large": 20897,
      "llama outperform": 18137,
      "low cost": 19268,
      "mitigate gap": 20252,
      "model exhibits": 20499,
      "making easily": 19502,
      "employs pretrained": 9264,
      "llava instructblip": 18245,
      "systematic exploration": 31872,
      "100k tokens": 51,
      "window size": 35613,
      "summaries despite": 31606,
      "importance task": 14192,
      "public llms": 26839,
      "common types": 5019,
      "develop automatic": 7910,
      "evaluation costs": 9934,
      "syntactically correct": 31830,
      "task translating": 32203,
      "language query": 16810,
      "explore design": 10579,
      "making llama": 19510,
      "possess capability": 25202,
      "advancements recent": 1475,
      "training recipe": 33596,
      "visual data": 35328,
      "diversity human": 8479,
      "exploiting large": 10559,
      "coding ability": 4865,
      "task given": 32134,
      "given llm": 12753,
      "previous tasks": 25889,
      "solving complex": 30507,
      "code llama": 4772,
      "training minimal": 33566,
      "shown potential": 29904,
      "training approach": 33441,
      "ppo training": 25358,
      "communication patterns": 5051,
      "accuracy results": 909,
      "backward reasoning": 3071,
      "question explored": 27068,
      "information paper": 14896,
      "paper formally": 24054,
      "formally define": 11752,
      "findings significant": 11255,
      "given problem": 12761,
      "method resulting": 19970,
      "resulting substantial": 28564,
      "llms standard": 18968,
      "consists key": 5764,
      "key modules": 15779,
      "learning module": 17612,
      "powerful text": 25352,
      "applying real": 2223,
      "presents significant": 25596,
      "biases research": 3685,
      "llm framework": 18311,
      "blind reviews": 3773,
      "completeness relevance": 5256,
      "value llms": 35024,
      "propose methodology": 26529,
      "mbert mt5": 19711,
      "languages limited": 16891,
      "reasoning visual": 27464,
      "extract text": 10746,
      "information required": 14904,
      "method pretrained": 19957,
      "endtoend approach": 9433,
      "pipeline approach": 24963,
      "helpfulness harmlessness": 13519,
      "efficient empirical": 9033,
      "resources compared": 28431,
      "process large": 26069,
      "method boosts": 19886,
      "method finetune": 19921,
      "finetune opensource": 11297,
      "use code": 34519,
      "highquality datasets": 13686,
      "language code": 16051,
      "approach approach": 2238,
      "models family": 21262,
      "information provided": 14900,
      "new form": 22804,
      "suggest reasoning": 31579,
      "capturing complex": 4156,
      "following human": 11691,
      "alignment simple": 1778,
      "weights pretrained": 35511,
      "models weights": 22126,
      "approach extend": 2277,
      "models chat": 21052,
      "results underscore": 28701,
      "effectiveness wide": 8975,
      "wide applicability": 35546,
      "video demonstrations": 35264,
      "methods generative": 20042,
      "text andor": 32817,
      "script generation": 29295,
      "based demonstration": 3151,
      "llms helpful": 18668,
      "users control": 34684,
      "rlhf stage": 28909,
      "datasets generates": 7125,
      "baselines trained": 3273,
      "digital age": 8189,
      "techniques particularly": 32655,
      "stateoftheart framework": 30931,
      "query resolution": 27031,
      "intricate tasks": 15485,
      "collection online": 4932,
      "opensourced code": 23549,
      "improved controllability": 14309,
      "direction finetuning": 8223,
      "lms prompting": 19103,
      "backbone lms": 3054,
      "gpt4 leads": 13088,
      "diverse finetuning": 8428,
      "findings regarding": 11247,
      "provides initial": 26756,
      "initial set": 14965,
      "questions language": 27116,
      "containing specific": 5836,
      "improve abilities": 14253,
      "methodology leverages": 19995,
      "derived pretrained": 7674,
      "llama2chat model": 18219,
      "model level": 20610,
      "llms garnered": 18628,
      "garnered considerable": 12120,
      "initial tokens": 14967,
      "additionally introduce": 1290,
      "specialized data": 30668,
      "results realworld": 28667,
      "training fewshot": 33520,
      "code generated": 4752,
      "llms susceptible": 18987,
      "engineering tactics": 9469,
      "reduce errors": 27709,
      "finally discuss": 11192,
      "code language": 4769,
      "models optimized": 21780,
      "models master": 21701,
      "text code": 32829,
      "popular entities": 25116,
      "works pretrained": 35821,
      "reranking generated": 28280,
      "poses challenging": 25168,
      "finegrained multimodal": 11275,
      "overcome challenges": 23920,
      "capability leveraging": 4097,
      "feature extraction": 11023,
      "consistent patterns": 5740,
      "including question": 14514,
      "accuracy respectively": 908,
      "methods highly": 20044,
      "uncovering hidden": 34137,
      "despite commendable": 7771,
      "meticulously crafted": 20115,
      "enhancing depth": 9560,
      "renowned datasets": 28083,
      "gpt4 series": 13116,
      "available instructiontuning": 2983,
      "detailed responses": 7841,
      "challenges introducing": 4353,
      "solution designed": 30473,
      "designed automatically": 7724,
      "generating instructions": 12431,
      "conversations chatgpt": 6106,
      "diversity number": 8481,
      "problem present": 26002,
      "sparsity different": 30629,
      "dataset including": 7002,
      "hard negative": 13422,
      "improved capability": 14306,
      "object hallucination": 23200,
      "positional bias": 25186,
      "bias use": 3666,
      "prompt produce": 26340,
      "input prompt": 15021,
      "random perturbations": 27176,
      "language interaction": 16099,
      "uses large": 34712,
      "harnessing large": 13463,
      "support large": 31709,
      "substantial promise": 31475,
      "health care": 13484,
      "albeit relatively": 1694,
      "model palm": 20680,
      "responses llms": 28499,
      "text retrieval": 32934,
      "handle longer": 13408,
      "obviating need": 23265,
      "categories extensive": 4223,
      "like openflamingo": 17890,
      "models incorporating": 21386,
      "understanding code": 34215,
      "thousands words": 33087,
      "hundreds thousands": 13947,
      "touvron et": 33304,
      "effectively handle": 8919,
      "tasks widely": 32551,
      "suboptimal quality": 31440,
      "effectively leverage": 8923,
      "paper available": 24016,
      "llms help": 18667,
      "capacity llms": 4127,
      "capabilities approach": 4001,
      "tasks enabling": 32309,
      "wider range": 35585,
      "training based": 33442,
      "efficiency llm": 9008,
      "adversely affecting": 1527,
      "finetuning opensource": 11466,
      "replacing entities": 28106,
      "pretraining llms": 25817,
      "like medicine": 17887,
      "technique based": 32617,
      "accuracy 92": 860,
      "requiring minimal": 28274,
      "study pretrained": 31378,
      "make predictions": 19478,
      "tuning learning": 33991,
      "vs llama": 35401,
      "llms match": 18795,
      "match surpass": 19647,
      "instructiontuning methods": 15308,
      "contributions opensource": 6041,
      "systematically evaluating": 31880,
      "bilingual evaluation": 3710,
      "evaluation understudy": 10023,
      "evaluation rouge": 10004,
      "according experiment": 846,
      "domain work": 8604,
      "models alignment": 20973,
      "identify important": 14009,
      "simple implement": 30154,
      "lines code": 18001,
      "larger batch": 17316,
      "increase throughput": 14607,
      "pretraining code": 25787,
      "model suite": 20813,
      "code replicate": 4809,
      "techniques fall": 32638,
      "works propose": 35823,
      "fundamental limitations": 11978,
      "searches efficient": 29315,
      "solve single": 30496,
      "method achieve": 19867,
      "level accuracy": 17730,
      "llama approach": 18075,
      "numerous downstream": 23183,
      "development nlp": 7964,
      "obtained finetuning": 23256,
      "promising technique": 26299,
      "information using": 14924,
      "fail produce": 10905,
      "benchmark model": 3398,
      "adapts pretrained": 1226,
      "text specific": 32948,
      "parametric knowledge": 24297,
      "number retrieved": 23159,
      "tokens generating": 33230,
      "textual description": 33026,
      "extraction methods": 10769,
      "potential using": 25305,
      "gpt4 opensource": 13102,
      "average f1score": 3017,
      "safe reinforcement": 29030,
      "regarding helpfulness": 27810,
      "cost models": 6251,
      "llms optimization": 18833,
      "method solve": 19975,
      "harmful responses": 13443,
      "llms open": 18826,
      "llms central": 18460,
      "tool utilization": 33261,
      "llms compromising": 18482,
      "compromising general": 5440,
      "general abilities": 12155,
      "method enhance": 19912,
      "capabilities open": 4055,
      "encoder layers": 9350,
      "trained solely": 33427,
      "data surprisingly": 6885,
      "previously overlooked": 25899,
      "directly process": 8240,
      "associated language": 2645,
      "prompts inputs": 26425,
      "visual recognition": 35353,
      "effectiveness pretrained": 8963,
      "visual encoding": 35330,
      "simple baseline": 30142,
      "approaches method": 2384,
      "require access": 28210,
      "encoder model": 9352,
      "images texts": 14089,
      "model ability": 20337,
      "ability extract": 665,
      "features input": 11033,
      "attention patterns": 2731,
      "llama simple": 18145,
      "instructionfollowing abilities": 15222,
      "domainspecific llms": 8653,
      "affecting performance": 1542,
      "provide public": 26723,
      "gap presenting": 12103,
      "demonstrate gpt2": 7459,
      "corresponding predictions": 6230,
      "various model": 35119,
      "datasets highlight": 7128,
      "ability outofdistribution": 707,
      "observe significant": 23233,
      "remarkable performances": 28051,
      "suboptimal results": 31441,
      "leading model": 17479,
      "hallucinations address": 13387,
      "encouraging model": 9402,
      "incorrect answers": 14582,
      "small fraction": 30343,
      "data response": 6841,
      "generation sota": 12602,
      "chatgpt large": 4482,
      "single turn": 30227,
      "requires users": 28266,
      "establish strong": 9770,
      "experiments effectiveness": 10438,
      "plays significant": 25031,
      "issue mainly": 15656,
      "llm particular": 18347,
      "instruct tuning": 15131,
      "inject domain": 14979,
      "suitable dataset": 31597,
      "notable performance": 23027,
      "models higher": 21354,
      "code work": 4830,
      "hours training": 13767,
      "final result": 11183,
      "chat benchmarks": 4441,
      "personalized generative": 24890,
      "efficiency code": 8997,
      "approach enhancing": 2272,
      "offering flexible": 23297,
      "leveraging recent": 17794,
      "methods training": 20104,
      "potential multimodal": 25280,
      "performance limitations": 24658,
      "dynamic field": 8759,
      "novel methodology": 23098,
      "language features": 16075,
      "optimizing various": 23658,
      "various facets": 35092,
      "including improved": 14495,
      "training respectively": 33600,
      "performance chinese": 24542,
      "spur future": 30832,
      "make fewer": 19469,
      "effect adding": 8850,
      "hallucinations challenging": 13389,
      "nature information": 22590,
      "based provided": 3213,
      "curated instructions": 6472,
      "comparing llms": 5192,
      "findings align": 11230,
      "dataset serves": 7036,
      "difficulty task": 8178,
      "evaluate use": 9868,
      "dataset synthetic": 7043,
      "models hallucination": 21345,
      "domain nlp": 8580,
      "requires highquality": 28254,
      "task focus": 32125,
      "image audio": 14055,
      "sequence text": 29609,
      "taking step": 32038,
      "specialized knowledge": 30673,
      "model enables": 20486,
      "approach producing": 2328,
      "gpt35turbo using": 13046,
      "useful various": 34642,
      "promise large": 26276,
      "languages bangla": 16864,
      "performance effect": 24579,
      "effect scaling": 8856,
      "work identify": 35720,
      "factors affecting": 10868,
      "potential performance": 25288,
      "computationally intensive": 5488,
      "data resolve": 6840,
      "derive simple": 7671,
      "manipulation framework": 19544,
      "opensource vlms": 23546,
      "comprehension models": 5346,
      "behavior different": 3312,
      "model gptj": 20561,
      "learningbased approach": 17674,
      "data observe": 6786,
      "varies widely": 35054,
      "2023 demonstrated": 233,
      "icl ability": 13974,
      "llms enabling": 18562,
      "incorrect predictions": 14587,
      "specific examples": 30693,
      "standard setting": 30881,
      "reveals current": 28814,
      "learning designed": 17557,
      "datasets languages": 7137,
      "quality human": 26967,
      "outputs code": 23887,
      "process challenging": 26052,
      "vast datasets": 35185,
      "linear programming": 17992,
      "current evaluation": 6488,
      "alignment method": 1771,
      "sacrificing performance": 29027,
      "codes checkpoints": 4850,
      "related questions": 27855,
      "space input": 30579,
      "inner product": 14988,
      "language structure": 16825,
      "pairs experiments": 23979,
      "widespread access": 35590,
      "models primarily": 21853,
      "models noteworthy": 21759,
      "language case": 16049,
      "alignment strategies": 1779,
      "video datasets": 35263,
      "understanding diverse": 34218,
      "finegrained perception": 11276,
      "overall framework": 23907,
      "learning use": 17667,
      "activate relevant": 1134,
      "news social": 22885,
      "evaluation introduce": 9962,
      "tasks assess": 32247,
      "bias models": 3658,
      "leading suboptimal": 17485,
      "dataset subset": 7041,
      "generation potential": 12572,
      "making models": 19512,
      "finetuning chatgpt": 11379,
      "pivotal role": 24980,
      "userfriendly interface": 34680,
      "increasing popularity": 14628,
      "investigating finetuning": 15608,
      "capability particularly": 4099,
      "study design": 31317,
      "direct responses": 8217,
      "finetuning especially": 11397,
      "news consumption": 22877,
      "work extend": 35708,
      "using carefully": 34745,
      "model utilize": 20855,
      "opendomain tasks": 23452,
      "need study": 22643,
      "execute complex": 10191,
      "task introduce": 32144,
      "llms combined": 18471,
      "paper seek": 24132,
      "specifically analyze": 30728,
      "applications various": 2182,
      "various fields": 35094,
      "effective use": 8905,
      "llms date": 18506,
      "support future": 31708,
      "speechbased slot": 30792,
      "proposed improve": 26602,
      "proposed integrate": 26603,
      "including gpt35turbo": 14487,
      "studies highlighted": 31271,
      "models solely": 21990,
      "solely focus": 30463,
      "preceding context": 25386,
      "entire context": 9623,
      "exhibits better": 10243,
      "designed mitigate": 7735,
      "task designed": 32109,
      "descriptions generate": 7686,
      "generate instructionfollowing": 12293,
      "derived image": 7673,
      "visual context": 35326,
      "trained realworld": 33421,
      "based proposed": 3212,
      "multimodal understanding": 22368,
      "models adapting": 20957,
      "distinct linguistic": 8366,
      "llm significantly": 18368,
      "information gain": 14870,
      "benchmark using": 3418,
      "diverse perspectives": 8447,
      "demographic groups": 7429,
      "express diverse": 10638,
      "certain groups": 4274,
      "datasets collected": 7076,
      "online reviews": 23368,
      "analysis common": 1913,
      "generation attracted": 12460,
      "paradigm instructiontuning": 24157,
      "neglecting potential": 22669,
      "student teacher": 31257,
      "distillation framework": 8338,
      "teacher student": 32585,
      "baselines zeroshot": 3275,
      "reliable systems": 27955,
      "required generate": 28229,
      "given partially": 12760,
      "time experiment": 33122,
      "algorithm enables": 1705,
      "possible model": 25215,
      "generated existing": 12355,
      "systems novel": 31909,
      "factual error": 10881,
      "investigation large": 15613,
      "llms marked": 18793,
      "usage memory": 34509,
      "multiagent environments": 22267,
      "enhancing llms": 9566,
      "highlighting significant": 13648,
      "methods face": 20033,
      "novel pipeline": 23104,
      "dataset case": 6947,
      "finetuning alpaca": 11371,
      "code implementation": 4766,
      "text distribution": 32846,
      "data distributions": 6680,
      "models degenerate": 21129,
      "model notably": 20658,
      "apply approach": 2203,
      "outputs finetuning": 23888,
      "tasks related": 32478,
      "model llmbased": 20627,
      "llmbased methods": 18389,
      "involves injecting": 15631,
      "reasoning information": 27412,
      "model consists": 20440,
      "sentences based": 29551,
      "propose search": 26564,
      "labels training": 15969,
      "results additionally": 28568,
      "end construct": 9409,
      "dataset variety": 7055,
      "tuning evaluating": 33976,
      "llms develop": 18537,
      "compared base": 5121,
      "model showing": 20779,
      "potentially causing": 25310,
      "improving robustness": 14421,
      "documents enabling": 8517,
      "aim explore": 1642,
      "languages investigate": 16882,
      "methods lora": 20065,
      "llama results": 18141,
      "llms chain": 18461,
      "human brain": 13799,
      "ability significantly": 721,
      "gpt4v llava": 13137,
      "intermediate representations": 15431,
      "construct benchmark": 5794,
      "methods achieving": 20002,
      "training instance": 33535,
      "automated manual": 2868,
      "programs written": 26205,
      "languages python": 16909,
      "accuracy various": 915,
      "types llama": 34066,
      "leads improved": 17494,
      "majority cases": 19446,
      "assessing models": 2612,
      "llms longer": 18784,
      "performance long": 24666,
      "techniques designed": 32634,
      "mixed results": 20270,
      "propose transform": 26578,
      "pretraining supervised": 25842,
      "unified simple": 34337,
      "inputoutput pair": 15039,
      "validate new": 34997,
      "llms notably": 18819,
      "modern large": 22161,
      "systems introduce": 31902,
      "models 3b": 20927,
      "3b parameters": 367,
      "projection layers": 26251,
      "respectively notably": 28463,
      "llms parameters": 18843,
      "icl changes": 13976,
      "llms hidden": 18669,
      "demonstrations overall": 7597,
      "work offers": 35738,
      "offers unique": 23314,
      "unique perspective": 34362,
      "capable answering": 4106,
      "enhance computational": 9509,
      "efficiency paper": 9009,
      "transformer training": 33743,
      "outperforming llms": 23799,
      "deployment resourceconstrained": 7659,
      "datasets larger": 7142,
      "work highlights": 35718,
      "generation roberta": 12598,
      "gptj models": 13151,
      "highlighted importance": 13640,
      "research reports": 28358,
      "accurate way": 929,
      "evaluation encompasses": 9944,
      "single linear": 30210,
      "existing video": 10323,
      "work serve": 35777,
      "practical scenarios": 25371,
      "aims analyze": 1658,
      "models extending": 21245,
      "llava model": 18248,
      "work builds": 35675,
      "captured existing": 4153,
      "pursuit artificial": 26889,
      "significant milestone": 29997,
      "perception understanding": 24464,
      "analyzing evaluating": 2002,
      "gap existing": 12089,
      "diffusion image": 8180,
      "prompt image": 26329,
      "introduce text": 15538,
      "dataset text": 7045,
      "recently experienced": 27597,
      "experienced rapid": 10373,
      "alignment capabilities": 1755,
      "strong alignment": 31161,
      "semantic comprehension": 29450,
      "finetuning lora": 11451,
      "small highquality": 30345,
      "dataset long": 7009,
      "alignment machine": 1769,
      "native language": 22499,
      "dataefficient alignment": 6928,
      "present automatic": 25514,
      "bert llama": 3517,
      "uncertainty estimates": 34117,
      "models original": 21781,
      "successful integration": 31534,
      "synthetic benchmark": 31846,
      "models testing": 22056,
      "llama2chat models": 18220,
      "finetuning similar": 11525,
      "opensource development": 23500,
      "models ai": 20971,
      "recognition textbased": 27642,
      "processes input": 26091,
      "enhancing overall": 9572,
      "creating efficient": 6366,
      "applications recent": 2174,
      "plms text": 25051,
      "widely explored": 35574,
      "encoderdecoder plms": 9372,
      "decoderonly llm": 7260,
      "strategy experimental": 31120,
      "computational burdens": 5456,
      "context token": 5922,
      "visual cues": 35327,
      "strategy significantly": 31128,
      "upper limit": 34484,
      "progress ai": 26208,
      "training introduce": 33536,
      "llms llama213b": 18779,
      "prompt types": 26356,
      "truefalse questions": 33927,
      "explicit implicit": 10545,
      "llm respectively": 18360,
      "effectively align": 8910,
      "increasing demand": 14621,
      "llmbased chatbots": 18386,
      "diffusion xl": 8185,
      "experiments validate": 10497,
      "average success": 3025,
      "rate base": 27263,
      "different instruction": 8084,
      "realworld deployment": 27339,
      "task work": 32209,
      "long prompt": 19175,
      "context short": 5917,
      "performance generalpurpose": 24614,
      "gpt35turbo training": 13045,
      "evaluation t5": 10017,
      "prompt settings": 26344,
      "baseline code": 3244,
      "reveal finetuned": 28797,
      "llms surpass": 18985,
      "points exact": 25073,
      "models encounter": 21202,
      "challenges field": 4344,
      "performance conditional": 24558,
      "employing lora": 9261,
      "achieving highest": 1093,
      "direction enhancing": 8222,
      "robust zeroshot": 28940,
      "opensource counterparts": 23495,
      "models persists": 21814,
      "code necessary": 4793,
      "necessary reproduce": 22608,
      "common style": 5016,
      "trainingfree method": 33650,
      "benchmark analysis": 3353,
      "increasingly recognized": 14644,
      "captions address": 4142,
      "visual prompts": 35348,
      "multiplechoice tasks": 22432,
      "findings based": 11231,
      "certain size": 4280,
      "logical thinking": 19158,
      "spatial localization": 30639,
      "automatically extracting": 2911,
      "addition provide": 1247,
      "given single": 12772,
      "single image": 30206,
      "additionally framework": 1288,
      "framework utilizes": 11904,
      "frozen large": 11936,
      "datasets evaluation": 7105,
      "studies primarily": 31278,
      "field challenges": 11134,
      "mainly relies": 19407,
      "showcasing immense": 29850,
      "content online": 5866,
      "online inference": 23364,
      "approach relies": 2334,
      "achieve reasonable": 980,
      "equipped efficient": 9681,
      "variable names": 35034,
      "toolaugmented llms": 33263,
      "enabling direct": 9318,
      "leverages knowledge": 17767,
      "opensource pretrained": 23540,
      "enabling arbitrary": 9316,
      "data serve": 6858,
      "matches exceeds": 19650,
      "generation integration": 12526,
      "original clip": 23700,
      "ability specifically": 723,
      "new document": 22793,
      "unlike traditional": 34405,
      "bert classification": 3500,
      "models grasp": 21339,
      "diverse attributes": 8413,
      "properties flexibility": 26472,
      "methods various": 20110,
      "achieving significantly": 1101,
      "work results": 35774,
      "opensourced model": 23555,
      "technique finetuning": 32622,
      "35 gpt": 340,
      "provide comparative": 26689,
      "outperforms established": 23818,
      "robust performance": 28938,
      "effective tool": 8902,
      "alignment aligning": 1753,
      "effectiveness llm": 8956,
      "methods paper": 20071,
      "datasets investigate": 7135,
      "modeling analysis": 20887,
      "alpaca7b model": 1836,
      "key metric": 15776,
      "metric evaluating": 20121,
      "prominent models": 26271,
      "pivotal insights": 24978,
      "reasoning enhanced": 27406,
      "prominent method": 26270,
      "algorithm achieves": 1700,
      "moving average": 22240,
      "average ema": 3014,
      "research aim": 28287,
      "incorporating external": 14575,
      "knowledge used": 15918,
      "methods limitations": 20061,
      "approaches straightforwardly": 2396,
      "commonly known": 5026,
      "known hallucination": 15934,
      "position encoding": 25183,
      "boosting llm": 3823,
      "pruning large": 26810,
      "datasets achieving": 7059,
      "compatible existing": 5204,
      "model vlm": 20861,
      "art model": 2515,
      "model codes": 20426,
      "enables dynamic": 9296,
      "examples behavior": 10118,
      "execution accuracy": 10198,
      "exhibit high": 10219,
      "accuracy relevance": 907,
      "strong capabilities": 31165,
      "techniques reinforcement": 32659,
      "models behave": 21010,
      "ways difficult": 35450,
      "model test": 20828,
      "naive finetuning": 22478,
      "work simple": 35787,
      "understanding identifying": 34231,
      "pull requests": 26877,
      "fostering collaboration": 11788,
      "nature software": 22591,
      "software projects": 30458,
      "communication channels": 5049,
      "explore zeroshot": 10606,
      "available models": 2989,
      "using closedsource": 34751,
      "7b13b 70b": 544,
      "model enhance": 20490,
      "dataset utilizing": 7054,
      "data conditions": 6651,
      "proprietary counterparts": 26635,
      "look leap": 19220,
      "challenging problems": 4394,
      "spanning domains": 30598,
      "coding task": 4868,
      "apply causal": 2204,
      "causal analysis": 4236,
      "models sizes": 21984,
      "125 million": 96,
      "requiring human": 28273,
      "single input": 30208,
      "sentences using": 29561,
      "models objective": 21762,
      "various strategies": 35148,
      "evaluation construct": 9933,
      "sampling techniques": 29098,
      "settings results": 29743,
      "impressive incontext": 14238,
      "shot learning": 29831,
      "pretrained capabilities": 25632,
      "current instruction": 6493,
      "ensuring data": 9611,
      "degrade model": 7379,
      "novel efficient": 23075,
      "candidate examples": 3992,
      "tasks deployment": 32291,
      "memory demands": 19812,
      "language built": 16045,
      "methods usually": 20108,
      "suffer significant": 31554,
      "significance llms": 29948,
      "llms utilizing": 19032,
      "utilizing external": 34970,
      "framework comprises": 11836,
      "parsing framework": 24320,
      "tasks determine": 32295,
      "tasks provide": 32464,
      "analysis effectively": 1919,
      "llms superior": 18982,
      "models vicuna7b": 22116,
      "evaluate llm": 9844,
      "llm robustness": 18364,
      "tendency hallucinate": 32712,
      "extensive research": 10711,
      "advantage unique": 1500,
      "struggle generating": 31241,
      "models codellms": 21059,
      "adapter module": 1202,
      "starcoder model": 30895,
      "code tokens": 4823,
      "novel datasets": 23072,
      "adverse effect": 1524,
      "llama generate": 18106,
      "lvlm llava": 19334,
      "images experimental": 14083,
      "development process": 7971,
      "used model": 34613,
      "unlike current": 34393,
      "automatic data": 2878,
      "decreased performance": 7300,
      "objective enhance": 23203,
      "model demonstrating": 20458,
      "mistral 7bs": 20230,
      "chatgpt gpt35turbo": 4477,
      "efficient llms": 9046,
      "additionally adaptive": 1270,
      "gains achieved": 12070,
      "rapid expansion": 27252,
      "offer opportunity": 23291,
      "models perspective": 21815,
      "present work": 25563,
      "training validation": 33643,
      "like falcon": 17861,
      "performance interpretability": 24635,
      "interpretability study": 15459,
      "content detection": 5856,
      "achieved humanlevel": 1010,
      "potential path": 25287,
      "feedback extensive": 11059,
      "human intervention": 13831,
      "model chatgpt": 20415,
      "unsupervised supervised": 34458,
      "demographic information": 7430,
      "inputs using": 15051,
      "given llms": 12754,
      "llms strong": 18970,
      "tokens input": 33234,
      "predict tokens": 25410,
      "performances variety": 24821,
      "context sizes": 5920,
      "t5 sequencetosequence": 31963,
      "models approaches": 20986,
      "code scripts": 4816,
      "language visual": 16855,
      "considerable computational": 5707,
      "addressing issues": 1375,
      "gpu cpu": 13169,
      "effective language": 8879,
      "devices work": 7991,
      "model gpt35": 20558,
      "long input": 19173,
      "leads significant": 17495,
      "benchmark method": 3397,
      "outperforming previous": 23802,
      "results accuracy": 28566,
      "augmenting llms": 2828,
      "synthesis capabilities": 31836,
      "text encoder": 32851,
      "performance making": 24671,
      "outperform original": 23781,
      "failures large": 10921,
      "assistance code": 2635,
      "chatgpt demonstrated": 4463,
      "llms suffer": 18979,
      "skills based": 30309,
      "vicuna guanaco": 35249,
      "rate 25": 27261,
      "examples incontext": 10127,
      "research computational": 28300,
      "computational framework": 5467,
      "greatly increased": 13273,
      "significantly limited": 30066,
      "highrisk setting": 13710,
      "novel computational": 23067,
      "13 different": 109,
      "higher degree": 13597,
      "ensure quality": 9605,
      "textual contexts": 33022,
      "strategically partitioning": 31099,
      "demand extensive": 7411,
      "availability highquality": 2963,
      "datasets remains": 7169,
      "imagetext tasks": 14095,
      "videotext tasks": 35270,
      "exhibit limitations": 10223,
      "traditional techniques": 33352,
      "problemsolving capabilities": 26039,
      "diverse problem": 8449,
      "problem scenarios": 26010,
      "scenarios extensive": 29204,
      "performance adaptability": 24516,
      "overcoming limitations": 23928,
      "instructions complete": 15247,
      "offline evaluation": 23321,
      "new online": 22826,
      "evaluation setting": 10010,
      "paper leverages": 24076,
      "improvement code": 14335,
      "datasets study": 7175,
      "findings lead": 11241,
      "discussion performance": 8300,
      "based transformers": 3233,
      "decrease general": 7297,
      "exhibits stateoftheart": 10252,
      "highlighting effectiveness": 13644,
      "datasets benchmarks": 7070,
      "lack proper": 15998,
      "advancement realm": 1458,
      "27b parameters": 285,
      "parameters effectively": 24242,
      "commendable performance": 4978,
      "enhance accuracy": 9503,
      "chatgpt 35": 4456,
      "offering promising": 23300,
      "sheer number": 29802,
      "retrieval specifically": 28754,
      "multiple images": 22395,
      "visual details": 35329,
      "relevant images": 27941,
      "image information": 14069,
      "range opensource": 27207,
      "opensource closedsource": 23488,
      "inputs based": 15043,
      "knowledge apply": 15812,
      "ai specifically": 1617,
      "llms offers": 18825,
      "offers promising": 23312,
      "retrieve information": 28767,
      "resistance hallucinations": 28397,
      "terms model": 32747,
      "model classification": 20420,
      "portuguese language": 25153,
      "stronger smaller": 31203,
      "benefit llms": 3482,
      "understanding query": 34262,
      "leverage models": 17758,
      "increase accuracy": 14592,
      "context results": 5914,
      "90 times": 592,
      "similar bert": 30097,
      "propose adversarial": 26493,
      "representation text": 28151,
      "unseen lowresource": 34441,
      "challenges machine": 4360,
      "languages previously": 16907,
      "models extensive": 21247,
      "translation llms": 33833,
      "prompt strategies": 26345,
      "finetuning crucial": 11383,
      "al 2023a": 1691,
      "xu et": 35884,
      "implementations available": 14165,
      "benefits use": 3492,
      "paper applies": 24012,
      "learning goals": 17573,
      "taxonomy automatically": 32575,
      "support claim": 31705,
      "set synthetic": 29707,
      "rules based": 29010,
      "various stages": 35146,
      "training compare": 33453,
      "llm architectures": 18270,
      "performance closedsource": 24545,
      "quality estimation": 26956,
      "correlates human": 6216,
      "candidate pool": 3993,
      "cases consistently": 4198,
      "varying numbers": 35178,
      "furthermore empirically": 11997,
      "scales linearly": 29156,
      "costly retraining": 6265,
      "retraining llms": 28727,
      "sensor data": 29525,
      "exhibits comparable": 10244,
      "studies highlight": 31270,
      "capability finetuned": 4087,
      "notably observe": 23032,
      "context prompts": 5909,
      "prompts significantly": 26440,
      "graph language": 13222,
      "text features": 32857,
      "knowledge distribution": 15837,
      "humans ability": 13917,
      "level applied": 17731,
      "sentences document": 29553,
      "instructions significantly": 15274,
      "primary cause": 25918,
      "ability address": 654,
      "llm designed": 18287,
      "providing insightful": 26774,
      "responses questions": 28509,
      "opensource algorithm": 23483,
      "finetuning multimodal": 11457,
      "users intentions": 34694,
      "following data": 11690,
      "ift datasets": 14036,
      "noticed models": 23042,
      "dataset potential": 7021,
      "establish new": 9769,
      "sizes notably": 30301,
      "previous opensource": 25871,
      "version gpt4": 35231,
      "improvement attributed": 14329,
      "problems using": 26034,
      "depends users": 7627,
      "particularly important": 24348,
      "important findings": 14202,
      "primarily studied": 25913,
      "experiments flant5": 10444,
      "data potential": 6800,
      "biases large": 3675,
      "significant promise": 30017,
      "applications including": 2156,
      "aim minimize": 1646,
      "biased outputs": 3670,
      "enhanced model": 9534,
      "reasoning fundamental": 27409,
      "transforms natural": 33806,
      "code prompts": 4800,
      "prompts elicit": 26413,
      "datasets conduct": 7081,
      "experiments understand": 10494,
      "understand code": 34189,
      "improve sample": 14298,
      "optimization large": 23628,
      "models gained": 21300,
      "gained immense": 12062,
      "importance recent": 14189,
      "remain unanswered": 27986,
      "context large": 5895,
      "improvements use": 14367,
      "connects models": 5690,
      "llms dynamic": 18550,
      "llms allows": 18430,
      "models advanced": 20967,
      "superior accuracy": 31644,
      "compute demands": 5494,
      "optimization including": 23627,
      "indicate potential": 14693,
      "potential challenges": 25248,
      "challenges model": 4361,
      "model hallucinations": 20563,
      "methods incorporating": 20050,
      "llms binary": 18451,
      "binary decision": 3734,
      "baseline method": 3252,
      "demonstrated high": 7525,
      "potential broader": 25246,
      "ai assistants": 1599,
      "different applications": 8046,
      "pivotal factor": 24977,
      "factor success": 10865,
      "current alignment": 6480,
      "harmlessness alignment": 13450,
      "harmless responses": 13448,
      "diminishes attack": 8203,
      "attack success": 2686,
      "rate asr": 27262,
      "harmful instructions": 13441,
      "jailbreak attacks": 15696,
      "attacks maintaining": 2693,
      "provide recommendations": 26726,
      "set recommendations": 29705,
      "methodology llms": 19996,
      "metrics use": 20149,
      "gain insight": 12058,
      "design task": 7715,
      "task better": 32087,
      "llms relatively": 18900,
      "framework identify": 11864,
      "significant bias": 29962,
      "contexts provide": 5946,
      "identify key": 14010,
      "llms analysis": 18431,
      "llms analyzing": 18432,
      "insights derived": 15072,
      "output correct": 23864,
      "fall categories": 10948,
      "model meets": 20637,
      "relatively large": 27887,
      "researchers limited": 28378,
      "current lvlms": 6512,
      "red teaming": 27690,
      "generate harmful": 12280,
      "terms different": 32743,
      "prominent opensourced": 26272,
      "common content": 5005,
      "sec filings": 29318,
      "capabilities consider": 4006,
      "cost latency": 6246,
      "data security": 6853,
      "security risk": 29350,
      "llama training": 18148,
      "including previous": 14513,
      "largescale llms": 17363,
      "prediction largescale": 25429,
      "multiple advanced": 22377,
      "advanced baselines": 1421,
      "set languages": 29691,
      "pretraining llama": 25816,
      "relying single": 27979,
      "methods present": 20076,
      "benchmark comprising": 3362,
      "tasks focused": 32342,
      "single multiple": 30215,
      "critical area": 6381,
      "need model": 22638,
      "education novel": 8836,
      "learning objective": 17623,
      "methods introduce": 20053,
      "distinguishing original": 8378,
      "various metrics": 35117,
      "additionally methods": 1292,
      "showing promising": 29863,
      "llm hallucinations": 18319,
      "rag systems": 27160,
      "evidence experiments": 10061,
      "reveal existing": 28796,
      "experts large": 10520,
      "scaling methods": 29174,
      "learning consequently": 17555,
      "changes time": 4412,
      "models matches": 21703,
      "knowledge rapidly": 15896,
      "diversity text": 8485,
      "making inefficient": 19505,
      "propose tokenlevel": 26575,
      "study multilingual": 31367,
      "detection models": 7873,
      "capabilities multimodal": 4046,
      "accurately interpreting": 937,
      "enhancing mllms": 9567,
      "systematic experiments": 31871,
      "maintains original": 19433,
      "benchmarks achieving": 3429,
      "facilitate exploration": 10839,
      "gpu 10": 13168,
      "pretrained context": 25638,
      "grouping using": 13302,
      "coherence creativity": 4888,
      "bloom 7b": 3783,
      "gptneo 13b": 13156,
      "times compared": 33158,
      "inference pretrained": 14799,
      "languages pretrained": 16904,
      "social situations": 30434,
      "executable code": 10189,
      "agents large": 1570,
      "used alternatives": 34583,
      "rely largescale": 27971,
      "various knowledgeintensive": 35104,
      "strategy improves": 31124,
      "generation strategy": 12605,
      "tasks enhance": 32312,
      "retrieval downstream": 28739,
      "tasks quality": 32467,
      "models vs": 22124,
      "hard understand": 13423,
      "relevant accurate": 27936,
      "results related": 28669,
      "aspects results": 2579,
      "reasoning different": 27404,
      "claude2 llama2": 4639,
      "properties llms": 26477,
      "degrees freedom": 7392,
      "valid solution": 34990,
      "increasingly used": 14647,
      "used various": 34636,
      "information textual": 14917,
      "model gpt4v": 20560,
      "study effects": 31323,
      "demonstrating initial": 7583,
      "generate desired": 12271,
      "presence hallucinations": 25509,
      "following similar": 11701,
      "designed implemented": 7733,
      "evaluation based": 9923,
      "gpt35turbo code": 13042,
      "efforts align": 9084,
      "access llm": 820,
      "dataset supervised": 7042,
      "novel technique": 23116,
      "complexity increases": 5303,
      "learn incontext": 17509,
      "query key": 27028,
      "weights input": 35507,
      "attention weight": 2743,
      "predicting word": 25419,
      "13b 30b": 119,
      "released github": 27923,
      "structured nature": 31225,
      "7b achieves": 532,
      "attributed key": 2760,
      "time constraints": 33113,
      "framework llms": 11881,
      "methods core": 20014,
      "llms select": 18930,
      "32 compared": 319,
      "inference compute": 14768,
      "operations large": 23571,
      "achieves perfect": 1055,
      "traditional models": 33349,
      "considerable efforts": 5708,
      "models black": 21030,
      "safeguard model": 29035,
      "model ownership": 20679,
      "predictions model": 25442,
      "dataset comprised": 6954,
      "1000 examples": 43,
      "generation humans": 12517,
      "llmbased systems": 18390,
      "tasks traditionally": 32533,
      "accuracy crucial": 866,
      "completion rate": 5262,
      "hallucinated answers": 13370,
      "predicted scores": 25415,
      "mistral llama": 20231,
      "loss llms": 19248,
      "depression detection": 7664,
      "novel paradigm": 23102,
      "method enhances": 19914,
      "phase models": 24917,
      "comprehensive approach": 5354,
      "showcasing potential": 29854,
      "efficacy method": 8988,
      "key techniques": 15787,
      "metric design": 20120,
      "longcontext llms": 19193,
      "metrics introduce": 20141,
      "services enhancing": 29661,
      "potential nlp": 25286,
      "llama2 aiming": 18165,
      "benefits process": 3491,
      "supervision large": 31697,
      "requires extensive": 28251,
      "errors using": 9729,
      "parameters family": 24244,
      "unleash potential": 34389,
      "base llms": 3122,
      "resources especially": 28434,
      "application multimodal": 2137,
      "comprehensive literature": 5385,
      "framework utilizing": 11905,
      "researchers conducted": 28372,
      "model findings": 20522,
      "models verifiable": 22113,
      "test accuracy": 32759,
      "common code": 5004,
      "despite significant": 7812,
      "implicit human": 14176,
      "addition present": 1244,
      "tasks proposed": 32463,
      "mt0 bloomz": 22253,
      "introduce extensive": 15507,
      "conduct detailed": 5597,
      "safety models": 29053,
      "capable solving": 4120,
      "thousands tokens": 33086,
      "training example": 33515,
      "llms common": 18473,
      "process generate": 26061,
      "execution evaluation": 10199,
      "design decisions": 7701,
      "understand factors": 34190,
      "framework evaluating": 11851,
      "training checkpoints": 33447,
      "cognitive bias": 4876,
      "cognitive biases": 4877,
      "biases llms": 3678,
      "varying effects": 35174,
      "bias mitigation": 3656,
      "proficient understanding": 26186,
      "understanding static": 34271,
      "lvlms suffer": 19338,
      "lvlms generate": 19336,
      "propose structured": 26570,
      "potential hallucination": 25257,
      "hallucination problems": 13382,
      "chatgpt vs": 4498,
      "stack overflow": 30855,
      "meta released": 19859,
      "study analyzing": 31300,
      "long term": 19184,
      "understand llms": 34195,
      "challenge human": 4314,
      "multiple large": 22398,
      "llms openais": 18829,
      "reallife applications": 27325,
      "methods dataset": 20017,
      "dataset 200": 6933,
      "better represent": 3624,
      "robust optimization": 28937,
      "llms witnessed": 19047,
      "llms face": 18598,
      "leading insufficient": 17477,
      "data point": 6796,
      "data apply": 6599,
      "llama 27b": 18061,
      "instructionbased prompting": 15216,
      "automatic summarization": 2897,
      "techniques address": 32626,
      "grand challenge": 13212,
      "various transformer": 35156,
      "significant increase": 29993,
      "exhibits notable": 10248,
      "problems varying": 26035,
      "mitigating hallucinations": 20263,
      "address hallucinations": 1329,
      "annotations work": 2033,
      "factuality generated": 10891,
      "additionally design": 1279,
      "algorithm proposed": 1712,
      "approach substantially": 2341,
      "substantially enhances": 31480,
      "accuracy llama": 894,
      "recently demonstrated": 27589,
      "drug discovery": 8748,
      "discovery process": 8276,
      "early results": 8780,
      "task provide": 32183,
      "allow efficient": 1801,
      "perception ability": 24462,
      "lays solid": 17458,
      "llmbased evaluation": 18387,
      "llm chatbots": 18280,
      "testing dataset": 32802,
      "gpt4 evaluation": 13073,
      "evaluation compared": 9930,
      "evaluation demonstrated": 9939,
      "evaluation llm": 9969,
      "validation future": 35005,
      "shown immense": 29885,
      "current largescale": 6506,
      "use opensource": 34558,
      "permissively licensed": 24866,
      "trained subset": 33429,
      "contexts adapting": 5936,
      "quantization model": 27010,
      "approaches results": 2392,
      "multilingual generalization": 22308,
      "domain datasets": 8560,
      "study address": 31294,
      "function calling": 11961,
      "increasingly prevalent": 14643,
      "13b parameter": 130,
      "llms basic": 18445,
      "llms inspired": 18722,
      "multiple programming": 22410,
      "accuracy numerical": 900,
      "relying solely": 27980,
      "language result": 16820,
      "suboptimal solutions": 31442,
      "overlook potential": 23945,
      "compared best": 5124,
      "finegrained annotations": 11269,
      "evaluating persona": 9911,
      "significant persona": 30007,
      "transformer attention": 33705,
      "predict correctness": 25406,
      "accurately detect": 933,
      "incorrect reasoning": 14588,
      "downstream accuracy": 8672,
      "models input": 21398,
      "accuracy llama2": 895,
      "llms wide": 19044,
      "prompting involves": 26380,
      "complex problem": 5282,
      "framework problem": 11890,
      "types evaluate": 34060,
      "data better": 6625,
      "language summaries": 16827,
      "instance level": 15109,
      "deployment hindered": 7647,
      "broader research": 3892,
      "specifically introduce": 30746,
      "128k context": 103,
      "gpt4 claude2": 13064,
      "inputs 100k": 15041,
      "scales llms": 29157,
      "potential superiority": 25300,
      "processing compared": 26098,
      "llama7b achieves": 18231,
      "humanannotated preference": 13886,
      "important problem": 14207,
      "use contrastive": 34521,
      "multimodal context": 22338,
      "features llms": 11034,
      "approaches bring": 2369,
      "llms numerous": 18821,
      "numerous new": 23187,
      "comparison gpt4": 5196,
      "different independent": 8082,
      "turbo gpt4": 34024,
      "ability accurately": 651,
      "models par": 21799,
      "shows opensource": 29931,
      "selection instruction": 29390,
      "emerges pivotal": 9187,
      "acquiring highquality": 1124,
      "unexplored research": 34302,
      "approaches llms": 2382,
      "approach inspired": 2300,
      "inspired observation": 15096,
      "challenging instructions": 4385,
      "evaluate difficulty": 9827,
      "challenging samples": 4396,
      "samples achieve": 29072,
      "generation explore": 12501,
      "respectively demonstrating": 28458,
      "lms demonstrate": 19079,
      "interactions increasingly": 15389,
      "using online": 34862,
      "generation enhance": 12491,
      "tuning llama2": 33993,
      "problem results": 26009,
      "lack indepth": 15993,
      "gpt4 largely": 13087,
      "various realworld": 35137,
      "billionscale llms": 3730,
      "challenges computational": 4338,
      "inference capabilities": 14763,
      "including roberta": 14517,
      "learning previous": 17632,
      "new annotation": 22775,
      "extra inference": 10739,
      "largescale experiments": 17352,
      "accuracy different": 868,
      "nonexistent objects": 22996,
      "popular mllms": 25128,
      "surprisingly simple": 31768,
      "answering openended": 2068,
      "introduce comprehensive": 15500,
      "8times faster": 586,
      "benchmark 15": 3350,
      "counterfactual examples": 6297,
      "examples propose": 10142,
      "particular identify": 24338,
      "capable text": 4121,
      "raising possibility": 27173,
      "llama demonstrated": 18091,
      "critical issue": 6388,
      "alignment study": 1780,
      "primarily focuses": 25911,
      "ai outputs": 1613,
      "expertise various": 10516,
      "llms generation": 18638,
      "advanced capabilities": 1422,
      "alignment algorithms": 1752,
      "align closely": 1725,
      "substantial potential": 31474,
      "multidocument question": 22277,
      "models type": 22087,
      "large visual": 17297,
      "medical imaging": 19769,
      "llms taken": 18990,
      "taken spotlight": 32026,
      "spotlight natural": 30822,
      "processing integrating": 26102,
      "vision enables": 35295,
      "users explore": 34687,
      "explore emergent": 10584,
      "vlms llava": 35378,
      "llava flamingo": 18243,
      "clip demonstrated": 4664,
      "various visiolinguistic": 35158,
      "visiolinguistic tasks": 35287,
      "consequently enormous": 5696,
      "enormous applications": 9586,
      "lack related": 15999,
      "better overall": 3613,
      "surged popularity": 31731,
      "performance visionlanguage": 24812,
      "essential component": 9757,
      "benchmark incorporating": 3393,
      "analysis spans": 1967,
      "various visionlanguage": 35162,
      "prediction uncertainty": 25439,
      "estimation approach": 9787,
      "models uncertainty": 22090,
      "importance measuring": 14188,
      "visual hallucinations": 35332,
      "hallucinations multimodal": 13397,
      "visual hallucination": 35331,
      "details image": 7845,
      "propose tool": 26576,
      "collect benchmark": 4912,
      "believe new": 3340,
      "hugging faces": 13785,
      "models huggingface": 21361,
      "size needed": 30265,
      "help avoid": 13503,
      "errors additionally": 9721,
      "multiple model": 22401,
      "following key": 11693,
      "pass1 metric": 24375,
      "tasks consider": 32279,
      "recover performance": 27667,
      "summary work": 31633,
      "approaches successfully": 2397,
      "steps model": 31059,
      "employs llms": 9263,
      "nonexpert individuals": 22998,
      "optimal hyperparameters": 23613,
      "key components": 15758,
      "datacentric approach": 6923,
      "resources provide": 28443,
      "provide explanations": 26698,
      "effect training": 8857,
      "data effectively": 6687,
      "reduces hallucinations": 27735,
      "automatic hallucination": 2883,
      "write code": 35846,
      "code large": 4770,
      "way large": 35440,
      "code instead": 4768,
      "using code": 34752,
      "feedback based": 11057,
      "humans finally": 13923,
      "introducing additional": 15550,
      "information results": 14907,
      "showcase effectiveness": 29836,
      "llama27bbased model": 18213,
      "llama270b model": 18202,
      "capabilities lvlms": 4043,
      "lvlms propose": 19337,
      "lower 50": 19283,
      "largely attributed": 17307,
      "work reveals": 35775,
      "stateoftheart lvlms": 30951,
      "propose multiple": 26533,
      "protein sequence": 26660,
      "high training": 13586,
      "growing trend": 13318,
      "limited research": 17961,
      "interactions centered": 15388,
      "opensource datasets": 23498,
      "finetuning enhance": 11395,
      "modeling domainspecific": 20890,
      "corpora given": 6166,
      "given rise": 12768,
      "papers primarily": 24148,
      "trained huge": 33402,
      "trillion parameters": 33905,
      "importantly model": 14218,
      "training compute": 33455,
      "analyze data": 1990,
      "constructed dataset": 5807,
      "strongest model": 31205,
      "code llm": 4776,
      "encounter difficulties": 9393,
      "provided data": 26738,
      "services context": 29660,
      "llms play": 18853,
      "jointly trains": 15716,
      "designed overcome": 7738,
      "demand computational": 7410,
      "ability leverage": 695,
      "efficient knowledge": 9042,
      "process translate": 26087,
      "llm process": 18351,
      "insight demonstrate": 15065,
      "challenges effectively": 4340,
      "novel perspective": 23103,
      "parameters generate": 24251,
      "advantages incontext": 1505,
      "internal mechanisms": 15438,
      "mechanisms models": 19754,
      "strongly biased": 31207,
      "llms learning": 18751,
      "task adaptation": 32074,
      "models users": 22100,
      "private data": 25958,
      "existing instruction": 10276,
      "text task": 32955,
      "reduces average": 27732,
      "capabilities paper": 4059,
      "generates token": 12405,
      "llms findings": 18608,
      "study offers": 31370,
      "llms contributing": 18493,
      "finegrained control": 11271,
      "user needs": 34662,
      "applications address": 2140,
      "method adopted": 19873,
      "enjoys better": 9579,
      "control llm": 6053,
      "llm generation": 18317,
      "tradeoff helpfulness": 33335,
      "redteaming large": 27695,
      "llm generates": 18315,
      "llms relying": 18903,
      "effective test": 8899,
      "connection problem": 5685,
      "coverage test": 6319,
      "toxic outputs": 33312,
      "relation hallucination": 27866,
      "standard instruction": 30877,
      "data addition": 6593,
      "inputs example": 15046,
      "vision transformers": 35310,
      "abilities solving": 641,
      "document paper": 8505,
      "enables llm": 9302,
      "iteratively refines": 15692,
      "quality finetuning": 26961,
      "ensemble method": 9597,
      "emerging large": 9193,
      "data varying": 6907,
      "considering diverse": 5724,
      "develop ensemble": 7916,
      "iteratively learn": 15691,
      "predict final": 25407,
      "theoretically optimal": 33053,
      "dataset method": 7010,
      "clickthrough rate": 4651,
      "diverse users": 8472,
      "key aspects": 15754,
      "propose targeted": 26571,
      "captioning address": 4136,
      "mllms recently": 20299,
      "immense popularity": 14114,
      "various ways": 35163,
      "like data": 17859,
      "novel data": 23070,
      "rigorous quality": 28884,
      "models ablation": 20935,
      "datasets annotated": 7065,
      "learn follow": 17506,
      "extraction datasets": 10765,
      "samples training": 29088,
      "performance surpasses": 24775,
      "data llm": 6758,
      "standardized benchmark": 30887,
      "scaling instruction": 29163,
      "scalable method": 29126,
      "inspired cognitive": 15094,
      "problem perspective": 26001,
      "combination low": 4951,
      "effectively mitigate": 8924,
      "term new": 32731,
      "importantly training": 14219,
      "20 training": 215,
      "legal documents": 17695,
      "domain furthermore": 8565,
      "finetuning paradigm": 11469,
      "selecting best": 29386,
      "open challenge": 23385,
      "selection approach": 29389,
      "approach avoids": 2241,
      "capabilities enabling": 4012,
      "leveraging taskspecific": 17795,
      "taskspecific demonstrations": 32561,
      "taskspecific examples": 32562,
      "resource requirements": 28415,
      "resourceefficient manner": 28425,
      "recommendation reasoning": 27650,
      "tasks respectively": 32487,
      "tasks tasks": 32522,
      "conduct empirical": 5598,
      "aforementioned models": 1553,
      "suggest potential": 31577,
      "study provide": 31384,
      "specialized applications": 30665,
      "yi model": 35906,
      "context models": 5902,
      "models deliver": 21130,
      "platforms like": 25011,
      "efforts pretraining": 9092,
      "data deduplication": 6670,
      "given current": 12745,
      "common language": 5008,
      "impressive accuracy": 14230,
      "capabilities notably": 4054,
      "sft data": 29758,
      "reliability generating": 27950,
      "scarcity publicly": 29195,
      "real data": 27310,
      "twostep approach": 34049,
      "cause analysis": 4253,
      "approach use": 2350,
      "ablation experiments": 734,
      "unified information": 34329,
      "slight decrease": 30321,
      "designed realworld": 7742,
      "understanding applications": 34209,
      "relatively low": 27889,
      "stateoftheart competitive": 30926,
      "accuracy scores": 911,
      "method offers": 19950,
      "offers valuable": 23317,
      "guiding future": 13359,
      "preprocessed dataset": 25504,
      "finetuned opensource": 11348,
      "using quantitative": 34894,
      "length limited": 17708,
      "address unique": 1359,
      "text lengths": 32909,
      "systems crucial": 31891,
      "based importance": 3174,
      "seamlessly integrate": 29304,
      "performance reducing": 24735,
      "computational time": 5483,
      "improvement generating": 14339,
      "parameter opensource": 24193,
      "adapting plms": 1214,
      "restricting use": 28534,
      "llama llava": 18122,
      "llms low": 18786,
      "perform data": 24482,
      "attention computation": 2715,
      "suggesting need": 31583,
      "designed optimize": 7737,
      "optimize computational": 23641,
      "patterns early": 24409,
      "early layers": 8778,
      "7bparameter model": 547,
      "model maintaining": 20633,
      "knowledge relevant": 15901,
      "certain scenarios": 4279,
      "llm makes": 18337,
      "llm achieving": 18263,
      "deployment large": 7651,
      "variety different": 35058,
      "models vllms": 22120,
      "everyday objects": 10057,
      "gemini llama2": 12139,
      "mistral models": 20234,
      "using newly": 34858,
      "collected corpus": 4920,
      "prompt work": 26357,
      "increased performance": 14611,
      "majority existing": 19447,
      "respectively experiments": 28459,
      "75 compared": 516,
      "scalable data": 29124,
      "training trajectories": 33639,
      "challenges complexity": 4337,
      "complexity finetuning": 5302,
      "data bridge": 6629,
      "dataset performance": 7020,
      "stateoftheart data": 30928,
      "50k data": 428,
      "accuracy challenging": 863,
      "al 2023b": 1692,
      "cost data": 6243,
      "algorithms language": 1719,
      "alignment phase": 1776,
      "strategy automatically": 31114,
      "ability execute": 664,
      "multiple sequential": 22418,
      "instructions existing": 15251,
      "response introduce": 28478,
      "dataset million": 7011,
      "using attention": 34734,
      "data engineering": 6691,
      "model attains": 20378,
      "run single": 29013,
      "single v100": 30229,
      "enabling llms": 9324,
      "llms traditional": 19005,
      "representation llms": 28144,
      "chatgpt35 tasks": 4502,
      "present generative": 25534,
      "consists components": 5762,
      "parallel training": 24172,
      "generation strategies": 12604,
      "novel solution": 23109,
      "access sensitive": 826,
      "challenge study": 4329,
      "llm create": 18284,
      "chainofthought approach": 4296,
      "llms question": 18883,
      "transfer llms": 33678,
      "comprehensive list": 5384,
      "vision foundation": 35300,
      "models combine": 21063,
      "models facilitating": 21256,
      "key features": 15767,
      "llm approach": 18269,
      "notably model": 23031,
      "4bit quantization": 413,
      "lora achieves": 19226,
      "tasks demonstrates": 32289,
      "adaptation capabilities": 1175,
      "easily available": 8793,
      "openais gpt35": 23445,
      "linguistic diversity": 18011,
      "settings original": 29740,
      "demonstrates substantial": 7575,
      "llama2 various": 18195,
      "greatly improves": 13272,
      "series empirical": 29634,
      "using 75": 34725,
      "important safetycritical": 14211,
      "safetycritical domains": 29056,
      "framework evaluate": 11850,
      "believe contributions": 3339,
      "peoples lives": 24453,
      "knowledge specifically": 15910,
      "27 billion": 282,
      "performing specific": 24838,
      "opening new": 23467,
      "llms mistral": 18800,
      "confidence accuracy": 5649,
      "recent ai": 27509,
      "progress achieving": 26207,
      "llms variety": 19033,
      "incorrect responses": 14589,
      "llama llms": 18123,
      "comprehend meaning": 5336,
      "effectively capture": 8914,
      "builds small": 3929,
      "design contrastive": 7700,
      "datasets proposed": 7162,
      "rtx 2080": 29000,
      "compared llava": 5146,
      "model shown": 20780,
      "proficiency generating": 26181,
      "highlights efficacy": 13651,
      "content various": 5876,
      "domain llm": 8573,
      "indicate flant5": 14688,
      "3b parameter": 366,
      "parameter llm": 24187,
      "llm embedding": 18295,
      "realm large": 27329,
      "significant expenses": 29981,
      "presents set": 25594,
      "set challenges": 29675,
      "dataset additionally": 6937,
      "text format": 32860,
      "finetuning previous": 11493,
      "importance using": 14194,
      "smaller sets": 30396,
      "method text": 19980,
      "machinegenerated texts": 19372,
      "methods tend": 20101,
      "lack explainability": 15988,
      "providing comprehensive": 26772,
      "criteria experimental": 6377,
      "achieving significant": 1100,
      "basis large": 3285,
      "recent explorations": 27520,
      "ratio high": 27275,
      "components image": 5312,
      "inference computation": 14765,
      "efficiently trained": 9075,
      "automatically produces": 2922,
      "produces diverse": 26160,
      "prompts resulting": 26438,
      "experiments involve": 10452,
      "corpus improve": 6184,
      "llama leveraging": 18119,
      "poses major": 25170,
      "risk assessment": 28893,
      "specific dataset": 30686,
      "key contributions": 15759,
      "deploying solutions": 7644,
      "generating effective": 12420,
      "designing data": 7752,
      "opensourced large": 23552,
      "llms crucial": 18500,
      "reasoning significantly": 27451,
      "distribution pretraining": 8395,
      "hallucinations based": 13388,
      "instructions technique": 15275,
      "focusing tasks": 11673,
      "tasks performed": 32450,
      "generalpurpose llm": 12254,
      "mllm benchmarks": 20293,
      "exploring state": 10622,
      "mechanism transformer": 19752,
      "transformer structure": 33740,
      "quadratic complexity": 26927,
      "fast inference": 10994,
      "linear scaling": 17997,
      "temporal dynamics": 32696,
      "winning recipe": 35617,
      "increasingly ubiquitous": 14646,
      "achieve propose": 979,
      "cot prompts": 6282,
      "exhibit enhanced": 10216,
      "addition conduct": 1237,
      "freeze parameters": 11918,
      "parameters llm": 24270,
      "achieved success": 1020,
      "novel blackbox": 23065,
      "model efficient": 20482,
      "wellknown transformer": 35527,
      "basic models": 3283,
      "faster speed": 11002,
      "hope proposed": 13755,
      "work step": 35788,
      "preserving model": 25609,
      "training diverse": 33501,
      "complex models": 5279,
      "assessing performance": 2613,
      "size similar": 30284,
      "bias analysis": 3642,
      "efficient large": 9043,
      "number input": 23145,
      "similar prior": 30114,
      "reduction approach": 27762,
      "model traditional": 20832,
      "provide important": 26706,
      "models field": 21268,
      "experiments prove": 10469,
      "llms opened": 18830,
      "performance representative": 24737,
      "data long": 6760,
      "insights propose": 15082,
      "appropriate prompts": 2406,
      "framework adapt": 11825,
      "considering high": 5725,
      "capture common": 4145,
      "adjust attention": 1390,
      "llms makes": 18792,
      "generated llm": 12372,
      "specific case": 30684,
      "providing flexibility": 26773,
      "social network": 30428,
      "content paper": 5867,
      "observe certain": 23226,
      "alignment generation": 1760,
      "right wrong": 28881,
      "llama fail": 18097,
      "degree language": 7387,
      "responses student": 28513,
      "repetitive tasks": 28095,
      "generative transformers": 12712,
      "nlp transformerbased": 22966,
      "models deal": 21125,
      "models context": 21104,
      "format accuracy": 11754,
      "compared widely": 5182,
      "combination finetuning": 4949,
      "answer directly": 2043,
      "new solutions": 22844,
      "generating captions": 12412,
      "performance widely": 24816,
      "systems usually": 31928,
      "performance previous": 24720,
      "recently witnessed": 27627,
      "methods significant": 20093,
      "automatic human evaluations": 2886,
      "sequence generation tasks": 29598,
      "large neural models": 17250,
      "mainly natural language": 19406,
      "efficacy pretrained checkpoints": 8990,
      "bert gpt2 roberta": 3511,
      "conducted extensive empirical": 5636,
      "extensive empirical study": 10684,
      "transformerbased models gpt2": 33763,
      "models gpt2 demonstrated": 21325,
      "language models automated": 16253,
      "contextualized word representations": 5964,
      "language models make": 16615,
      "general domain data": 12161,
      "automatic human evaluation": 2885,
      "language models existing": 16332,
      "require extensive human": 28215,
      "downstream nlp tasks": 8683,
      "wide range applications": 35550,
      "generation paper propose": 12568,
      "largescale pretrained models": 17378,
      "english language model": 9483,
      "outperforms existing baselines": 23820,
      "external knowledge bases": 10729,
      "83 billion parameter": 570,
      "parameter gpt2 model": 24183,
      "visual question answering": 35350,
      "model gpt2 generate": 20555,
      "stateoftheart text generators": 31000,
      "achieving impressive performance": 1096,
      "language paper propose": 16770,
      "results wide range": 28712,
      "tasks demonstrate effectiveness": 32287,
      "language modeling benchmarks": 16216,
      "language model results": 16198,
      "models era largescale": 21216,
      "language generation gpt2": 16084,
      "quality generated text": 26964,
      "performs better par": 24844,
      "better par stateoftheart": 3616,
      "text generation proposed": 32881,
      "code facilitate future": 4751,
      "increase model complexity": 14599,
      "transformerbased unidirectional language": 33771,
      "machine learning approaches": 19344,
      "bert openai gpt2": 3522,
      "evaluate results using": 9864,
      "results using rouge": 28703,
      "generation using pretrained": 12635,
      "social media provide": 30427,
      "data paper propose": 6793,
      "language model automatically": 16121,
      "improve quality generated": 14291,
      "finetuning large pretrained": 11433,
      "language models capture": 16273,
      "language models largescale": 16410,
      "models largescale language": 21433,
      "text various domains": 32964,
      "simple effective method": 30145,
      "approach significantly improves": 2338,
      "language models text": 16730,
      "text corpus used": 32837,
      "natural language modeling": 22532,
      "model generates valid": 20547,
      "analyze impact different": 1995,
      "methods require finetuning": 20088,
      "work investigate use": 35728,
      "investigate use pretrained": 15599,
      "language models propose": 16670,
      "data models code": 6778,
      "gpt2 pretrained model": 12940,
      "language model new": 16180,
      "text generation important": 32870,
      "generative models suffer": 12679,
      "generation models based": 12553,
      "models based gpt2": 21008,
      "based gpt2 model": 3169,
      "paper present new": 24090,
      "data augmentation finetuning": 6608,
      "text generation language": 32871,
      "human preferences results": 13857,
      "recently deep generative": 27588,
      "propose novel model": 26552,
      "generation challenging task": 12471,
      "largescale language model": 17359,
      "metrics human evaluation": 20140,
      "generate semantically correct": 12322,
      "human evaluation study": 13809,
      "achieve better results": 949,
      "paper propose unsupervised": 24118,
      "current stateoftheart models": 6536,
      "language model evaluate": 16137,
      "generation models generate": 12554,
      "generation model gpt2": 12551,
      "transformer architectures models": 33704,
      "generate natural language": 12305,
      "natural language captions": 22509,
      "adapt language model": 1165,
      "paper presents novel": 24103,
      "presents novel approach": 25589,
      "unidirectional language models": 34322,
      "artificially generated texts": 2543,
      "paper explore use": 24048,
      "generative model gpt2": 12673,
      "evaluation results method": 10002,
      "results method achieves": 28644,
      "models text generation": 22058,
      "models generated text": 21312,
      "mental health support": 19841,
      "recent studies shown": 27556,
      "human evaluation demonstrate": 13806,
      "architecture language modeling": 2446,
      "conditional text generation": 5572,
      "improving language understanding": 14412,
      "data finetuned gpt2": 6714,
      "lack training data": 16007,
      "generative pretraining gpt2": 12702,
      "significantly outperforms baselines": 30075,
      "address issues introduce": 1337,
      "large pretrained transformer": 17268,
      "provided natural language": 26742,
      "bias language models": 3649,
      "language models predicting": 16659,
      "long document summarization": 19171,
      "low resource setting": 19276,
      "problem proposing novel": 26007,
      "experiments various datasets": 10499,
      "datasets natural language": 7153,
      "models including bert": 21379,
      "including bert roberta": 14462,
      "bert roberta t5": 3532,
      "tasks main categories": 32411,
      "model sizes data": 20797,
      "best performance single": 3569,
      "results experimental results": 28609,
      "experimental results proposed": 10405,
      "results proposed approach": 28660,
      "benchmark natural language": 3403,
      "recently increasing number": 27602,
      "unified evaluation framework": 34326,
      "language models identify": 16380,
      "offtheshelf language models": 23328,
      "adopt curriculum learning": 1402,
      "effectiveness proposed method": 8965,
      "work pave way": 35743,
      "contextual word representations": 5957,
      "generation results indicate": 12597,
      "text training data": 32958,
      "training data code": 33470,
      "data code data": 6635,
      "stateoftheart results wide": 30990,
      "language modeling objectives": 16223,
      "larger model sizes": 17329,
      "language models experiments": 16334,
      "model answer questions": 20367,
      "models perform poorly": 21810,
      "language representation models": 16818,
      "using gpt2 model": 34790,
      "gpt2 model finetuned": 12918,
      "results showed finetuned": 28679,
      "showed finetuned model": 29857,
      "social biases study": 30419,
      "text generation methods": 32875,
      "propose new framework": 26538,
      "obtain better performance": 23248,
      "existing approaches rely": 10262,
      "transformer t5 model": 33742,
      "evaluation benchmarks method": 9927,
      "language models use": 16742,
      "overall results suggest": 23914,
      "transformers bert generative": 33778,
      "lms different architectures": 19083,
      "language models t5": 16725,
      "offensive toxic responses": 23285,
      "generation experimental results": 12497,
      "dataset demonstrate proposed": 6970,
      "approach significantly outperforms": 2339,
      "processing nlp recently": 26118,
      "use language models": 34541,
      "use transformer architecture": 34579,
      "experiments conducted benchmark": 10427,
      "downstream natural language": 8679,
      "shown promising performance": 29908,
      "deployment language models": 7650,
      "results language models": 28636,
      "language models significantly": 16705,
      "results proposed method": 28661,
      "models gpt3 t5": 21328,
      "general nlp tasks": 12181,
      "method achieves better": 19869,
      "comprehensive ablation studies": 5352,
      "knowledge distillation kd": 15832,
      "models propose novel": 21868,
      "does require finetuning": 8539,
      "text paper propose": 32915,
      "conducted extensive experiments": 5638,
      "address challenge paper": 1310,
      "pretraining language model": 25806,
      "method improves performance": 19931,
      "demonstrate model achieves": 7475,
      "model improves performance": 20575,
      "performance response generation": 24741,
      "widely adopted transformer": 35570,
      "simple effective approach": 30144,
      "prompt based method": 26309,
      "like bert gpt2": 17848,
      "gpt2 language modeling": 12910,
      "learning approach jointly": 17537,
      "address issue introduce": 1332,
      "training small number": 33617,
      "dialogue systems recent": 8021,
      "achieving stateoftheart performance": 1107,
      "stateoftheart performance various": 30976,
      "tasks finetuning pretrained": 32339,
      "language models used": 16743,
      "pretrained models finetuning": 25724,
      "pretrained transformer model": 25765,
      "different data sets": 8065,
      "performance variety tasks": 24796,
      "larger models compared": 17332,
      "model models trained": 20646,
      "number training data": 23170,
      "number trainable parameters": 23168,
      "achieves comparable better": 1037,
      "fewer trainable parameters": 11093,
      "provide useful insights": 26733,
      "visual textual modalities": 35358,
      "work shown large": 35782,
      "shown large language": 29894,
      "providing natural language": 26777,
      "improving task performance": 14424,
      "430 percentage points": 391,
      "contextualizing language models": 5967,
      "models despite success": 21148,
      "models pretrained natural": 21846,
      "pretrained natural language": 25736,
      "natural language data": 22511,
      "generate highquality short": 12286,
      "text generation propose": 32880,
      "test set best": 32785,
      "set best model": 29673,
      "based pretrained language": 3209,
      "models plms gpt2": 21821,
      "labeled training data": 15959,
      "training data lowresource": 33481,
      "decoderonly language model": 7256,
      "state art performance": 30909,
      "ability pretrained language": 713,
      "natural language prompts": 22561,
      "training data directly": 33474,
      "approach outperforms stateoftheart": 2324,
      "remains open question": 28007,
      "small set parameters": 30368,
      "way introduce new": 35439,
      "study propose novel": 31382,
      "language generation need": 16087,
      "generation need training": 12562,
      "overall quality generated": 23912,
      "language model similar": 16201,
      "spectrum natural language": 30773,
      "structured knowledge llms": 31224,
      "exact match score": 10090,
      "training data makes": 33482,
      "models deep language": 21127,
      "latent diffusion models": 17407,
      "model trained using": 20838,
      "models increasingly popular": 21392,
      "stateoftheart natural language": 30963,
      "generation models including": 12555,
      "language models prior": 16664,
      "various text generation": 35154,
      "available open source": 2993,
      "llms produce impressive": 18872,
      "language models offer": 16637,
      "area natural language": 2479,
      "ordinary differential equations": 23687,
      "large number trainable": 17255,
      "address limitations propose": 1346,
      "achieve stateoftheart results": 992,
      "domain pretrained language": 8584,
      "language models limited": 16420,
      "competitive performance zeroshot": 5228,
      "train large language": 33365,
      "models bert roberta": 21019,
      "pretrained models clip": 25721,
      "active research area": 1147,
      "llms requires expensive": 18912,
      "benchmark datasets using": 3374,
      "previous work shown": 25894,
      "size language models": 30254,
      "machine translation task": 19367,
      "50 fewer parameters": 422,
      "models achieved impressive": 20949,
      "huge model size": 13780,
      "language modeling capabilities": 16217,
      "generalization downstream tasks": 12212,
      "strong zeroshot performance": 31195,
      "pretrained language modelbased": 25663,
      "models propose new": 21866,
      "propose new paradigm": 26541,
      "knowledgeintensive nlp tasks": 15931,
      "zeroshot capabilities large": 35955,
      "evaluate performance models": 9857,
      "language model families": 16140,
      "benchmark dataset results": 3372,
      "language models detect": 16307,
      "machine translation systems": 19366,
      "little attention paper": 18050,
      "paper make attempt": 24078,
      "models paper introduce": 21793,
      "generation extensive experiments": 12504,
      "future research code": 12043,
      "future research field": 12047,
      "language models code": 16278,
      "tasks code generation": 32266,
      "code generation tasks": 4764,
      "generation tasks pretrained": 12616,
      "aligned human values": 1738,
      "nlp classification tasks": 22925,
      "evaluation human evaluation": 9960,
      "achieves superior results": 1078,
      "performance proposed model": 24725,
      "address shortcomings propose": 1358,
      "compared model finetuned": 5150,
      "using single nvidia": 34914,
      "prompt tuning prompt": 26353,
      "tuning prompt tuning": 34010,
      "training data prompt": 33488,
      "consistently outperforms existing": 5756,
      "outperforms existing models": 23821,
      "text autoregressive language": 32820,
      "languages experimental results": 16873,
      "significantly outperforms previous": 30078,
      "diverse set multimodal": 8460,
      "image captioning visual": 14058,
      "diffusion language model": 8182,
      "models work present": 22135,
      "language models generalize": 16357,
      "perform various tasks": 24508,
      "incontext learning performance": 14565,
      "propose reinforcement learning": 26562,
      "lack domain knowledge": 15985,
      "demonstrated superior performance": 7552,
      "original training data": 23724,
      "training data specifically": 33491,
      "data specifically propose": 6879,
      "language models knowledge": 16397,
      "models ranging 1b": 21890,
      "model families including": 20514,
      "language model propose": 16193,
      "reasoning capabilities smaller": 27389,
      "proved effective inducing": 26670,
      "abilities smaller models": 640,
      "models exhibit emergent": 21226,
      "shown impressive performance": 29888,
      "bert roberta bart": 3528,
      "models wide range": 22128,
      "achieving state art": 1105,
      "reasoning capabilities models": 27388,
      "arithmetic commonsense symbolic": 2501,
      "commonsense symbolic reasoning": 5045,
      "approach text generation": 2348,
      "comparable performance finetuned": 5086,
      "transformers large language": 33789,
      "models exhibit strong": 21228,
      "models plms t5": 21822,
      "large publicly available": 17273,
      "speech language models": 30783,
      "crossmodal representation alignment": 6425,
      "models trained public": 22071,
      "scale large language": 29138,
      "language model human": 16154,
      "results shed light": 28676,
      "language model instruction": 16157,
      "work shown finetuning": 35780,
      "shown finetuning large": 29878,
      "evaluation framework measure": 9952,
      "language model specifically": 16204,
      "gpt2 model pretrained": 12921,
      "massive text data": 19632,
      "recent years tasks": 27583,
      "models specific tasks": 21999,
      "outperform larger language": 23777,
      "language models learning": 16414,
      "conduct comprehensive ablation": 5590,
      "automatic quantitative evaluation": 2893,
      "similar large language": 30107,
      "language model fewshot": 16142,
      "shown excellent performance": 29874,
      "perform complex reasoning": 24477,
      "publicly available data": 26852,
      "generative pretrained models": 12691,
      "paper present comprehensive": 24088,
      "conduct thorough evaluation": 5627,
      "models existing works": 21232,
      "using constrained decoding": 34757,
      "clinical language models": 4660,
      "tasks remains unclear": 32481,
      "amazon mechanical turk": 1870,
      "large models like": 17236,
      "question answering vqa": 27061,
      "model based transformer": 20390,
      "models lms increasingly": 21674,
      "paving way future": 24422,
      "results proposed model": 28662,
      "demonstrate effectiveness proposed": 7449,
      "different target language": 8147,
      "models llms gpt35": 21546,
      "emerged promising approach": 9165,
      "outperforms stateoftheart methods": 23854,
      "work propose simple": 35759,
      "visual input experiments": 35335,
      "language model reward": 16199,
      "end propose novel": 9418,
      "models llms popular": 21601,
      "model code available": 20422,
      "using pretrained models": 34881,
      "different pretrained models": 8124,
      "recent proliferation large": 27546,
      "model works phases": 20873,
      "works phases phase": 35820,
      "policy optimization algorithm": 25087,
      "outperforms stateoftheart supervised": 23856,
      "gpt3 despite having": 12991,
      "efficient finetuning language": 9035,
      "llama 7b model": 18069,
      "sequence generation task": 29597,
      "extensive experimental evaluation": 10689,
      "approximation fisher information": 2420,
      "fisher information matrix": 11570,
      "success heavily relies": 31512,
      "models data released": 21122,
      "nlp tasks including": 22953,
      "approaches used training": 2401,
      "zero fewshot scenarios": 35942,
      "chatgpt gpt4 demonstrated": 4480,
      "knowledge work focus": 15921,
      "provide detailed analysis": 26694,
      "shown highly effective": 29884,
      "opt language model": 23595,
      "address issues present": 1338,
      "stateoftheart performance challenging": 30971,
      "opens new avenues": 23479,
      "used language model": 34607,
      "tasks prompt learning": 32460,
      "textual information news": 33031,
      "gptj llama models": 13149,
      "gpt2 models finetuned": 12926,
      "visual instruction tuning": 35340,
      "instruction tuning instruction": 15194,
      "tuning instruction tuning": 33984,
      "models llms using": 21658,
      "llms using machinegenerated": 19027,
      "using machinegenerated instructionfollowing": 34834,
      "zeroshot capabilities new": 35958,
      "capabilities new tasks": 4053,
      "paper present attempt": 24086,
      "present attempt use": 25513,
      "revolutionizing natural language": 28855,
      "models ability generate": 20934,
      "code model checkpoints": 4780,
      "code pretrained model": 4798,
      "foundation models fms": 11799,
      "transformers language models": 33787,
      "models llms showcased": 21627,
      "llms showcased remarkable": 18937,
      "model specifically designed": 20806,
      "alignment domainspecific instructions": 1757,
      "code models available": 4787,
      "based language models": 3184,
      "align large language": 1729,
      "remarkable performance diverse": 28045,
      "performance diverse domains": 24570,
      "simple highly effective": 30153,
      "output large language": 23870,
      "models llms human": 21558,
      "address challenges propose": 1316,
      "challenges propose novel": 4372,
      "vicuna large language": 35251,
      "problems paper propose": 26031,
      "additionally conduct comprehensive": 1276,
      "range nlp tasks": 27206,
      "llms demonstrated significant": 18525,
      "does require additional": 8538,
      "visual language models": 35343,
      "llms compared previous": 18476,
      "language model named": 16179,
      "instruction tuning make": 15202,
      "extensive experiments tasks": 10703,
      "existing pretrained language": 10309,
      "encoder visionlanguage models": 9359,
      "visionlanguage models vlms": 35323,
      "large visionlanguage models": 17292,
      "hundreds millions parameters": 13946,
      "introduce new paradigm": 15521,
      "task automatically generating": 32084,
      "substantial improvements compared": 31468,
      "models llms focusing": 21533,
      "open pretrained transformers": 23413,
      "pretrained transformers opt": 25773,
      "models finetuning pretrained": 21284,
      "task misinformation detection": 32157,
      "demonstrate efficiency effectiveness": 7453,
      "requires model learn": 28258,
      "llms chatgpt openai": 18467,
      "parameterefficient finetuning techniques": 24212,
      "using roberta t5": 34905,
      "make attempt investigate": 19452,
      "benchmarks demonstrate effectiveness": 3435,
      "extensive ablation studies": 10676,
      "llms human values": 18680,
      "capabilities work propose": 4085,
      "models reasoning capabilities": 21900,
      "terms zeroshot task": 32754,
      "zero fewshot prompting": 35940,
      "language models detecting": 16308,
      "detecting factual errors": 7856,
      "models llms explore": 21529,
      "incontext learning incontext": 14558,
      "learning incontext learning": 17587,
      "remains underexplored paper": 28019,
      "data used finetune": 6902,
      "alpaca experimental results": 1829,
      "language models scaling": 16699,
      "wide range topics": 35564,
      "llms significantly enhanced": 18950,
      "enhanced performance fewshot": 9536,
      "performs poorly context": 24853,
      "achieves performance levels": 1057,
      "compared previous stateoftheart": 5164,
      "models transformerbased pretrained": 22084,
      "bert gpt2 t5": 3512,
      "nlp tasks shown": 22963,
      "study provides valuable": 31386,
      "observed finetuned models": 23237,
      "decoderonly large language": 7258,
      "lower computational cost": 19286,
      "significant room improvement": 30025,
      "incontext learning performs": 14566,
      "models finetuned english": 21275,
      "human llm evaluations": 13843,
      "evaluating performance llms": 9910,
      "tasks demonstrate superiority": 32288,
      "end introduce new": 9413,
      "demonstrate competitive performance": 7442,
      "extremescale language models": 10806,
      "demonstrated exceptional performance": 7522,
      "language model efficiency": 16135,
      "gpt35 achieve similar": 13016,
      "different levels complexity": 8093,
      "abilities various domains": 647,
      "quality generated content": 26963,
      "paper study task": 24138,
      "language models visual": 16750,
      "perform complex tasks": 24478,
      "language models vicuna": 16748,
      "modern pretrained language": 22170,
      "social media data": 30424,
      "impact model performance": 14132,
      "widespread use language": 35598,
      "backpack language model": 3066,
      "results human evaluation": 28624,
      "code model data": 4781,
      "building better base": 3920,
      "better base models": 3593,
      "multitask learning approach": 22451,
      "information large language": 14878,
      "chatgpt gpt4 llama": 4481,
      "augmenting large language": 2826,
      "llms llama opt": 18772,
      "llms use tools": 19021,
      "model llm finetuned": 20622,
      "models lms typically": 21682,
      "llms gpt llama2": 18646,
      "based observations propose": 3205,
      "transfer learning approach": 33677,
      "recently shown promising": 27625,
      "instructiontuning language models": 15306,
      "based observation propose": 3203,
      "training data despite": 33473,
      "paper investigates effectiveness": 24073,
      "llms gpt4 gpt35": 18655,
      "language models prompt": 16668,
      "examples paper propose": 10140,
      "using parameterefficient finetuning": 34870,
      "unlike previous works": 34399,
      "models llms difficult": 21513,
      "truthfulness large language": 33941,
      "number attention heads": 23140,
      "models achieve higher": 20945,
      "natural language model": 22531,
      "humans language models": 13926,
      "trained llama 7b": 33408,
      "models evaluated human": 21220,
      "instruction tuning language": 15196,
      "models demonstrated ability": 21135,
      "settings findings reveal": 29735,
      "training data results": 33490,
      "largescale neural networks": 17370,
      "like large language": 17880,
      "models llms work": 21666,
      "models foundation models": 21294,
      "remains largely unexplored": 28003,
      "generated text findings": 12393,
      "model additional training": 20358,
      "binary classification tasks": 3733,
      "llms finetuning process": 18614,
      "domain large language": 8572,
      "llms including opensource": 18705,
      "instruction tuning approach": 15187,
      "general language models": 12173,
      "address challenges paper": 1315,
      "based gpt2 architecture": 3168,
      "tokens using novel": 33252,
      "investigate ability pretrained": 15572,
      "domains computer vision": 8616,
      "acquire general knowledge": 1117,
      "direction future research": 8225,
      "abstract meaning representation": 771,
      "meaning representation amr": 19722,
      "given text current": 12777,
      "data release code": 6832,
      "unlike previous work": 34398,
      "pretrained model better": 25713,
      "model extensive experiments": 20507,
      "transformer gpt models": 33719,
      "gpt3 gpt35 gpt4": 12997,
      "performance standard benchmarks": 24766,
      "achieve competitive results": 956,
      "domain recent advancements": 8588,
      "recent advancements language": 27491,
      "models lms led": 21675,
      "exceptional capabilities wide": 10168,
      "capabilities wide range": 4083,
      "based extensive experiments": 3158,
      "introduce simple effective": 15533,
      "remains relatively unexplored": 28011,
      "unexplored paper presents": 34301,
      "models llms based": 21484,
      "data paper present": 6792,
      "generate large number": 12300,
      "llms future research": 18624,
      "method does require": 19904,
      "compared supervised methods": 5176,
      "language models address": 16242,
      "finetuning peft techniques": 11477,
      "language models proposed": 16671,
      "integration large language": 15341,
      "textbased large language": 32974,
      "language models method": 16621,
      "demonstrate significant improvement": 7493,
      "language model multimodal": 16178,
      "curriculum learning strategy": 6549,
      "assess models performance": 2600,
      "opensource models llama": 23533,
      "comparable superior performance": 5093,
      "openai gpt2 model": 23436,
      "large lms llms": 17227,
      "demonstrated remarkable abilities": 7542,
      "llama open foundation": 18134,
      "nvidia a100 80gb": 23192,
      "tasks using various": 32546,
      "ablation studies investigate": 736,
      "tasks evaluate stateoftheart": 32317,
      "biomedical natural language": 3746,
      "models zero fewshot": 22141,
      "models better suited": 21024,
      "comprehensive evaluation multiple": 5373,
      "flant5 gpt35 gpt4": 11595,
      "tasks illustrating promising": 32358,
      "racial gender bias": 27146,
      "generative visionlanguage models": 12714,
      "hallucinations large language": 13392,
      "new benchmark dataset": 22781,
      "development safer reliable": 7974,
      "zero shot setting": 35946,
      "state art natural": 30906,
      "art natural language": 2517,
      "model specifically tuned": 20807,
      "results indicate models": 28633,
      "performance various downstream": 24799,
      "human evaluations finetuned": 13811,
      "framework training large": 11899,
      "visionlanguage models introduce": 35316,
      "language reasoning capabilities": 16813,
      "different model architectures": 8105,
      "presents effective approach": 25581,
      "models llm foundation": 21470,
      "llm foundation models": 18310,
      "models emergent capabilities": 21190,
      "evaluation metrics measure": 9976,
      "multimodal foundation models": 22343,
      "distillation large language": 8341,
      "given unique characteristics": 12782,
      "outperforms baselines various": 23809,
      "commercial opensource models": 4995,
      "datasets extensive evaluation": 7115,
      "wide range downstream": 35552,
      "substantial parameter size": 31472,
      "enhanced reasoning capabilities": 9539,
      "chainofthought cot prompting": 4298,
      "models knowledge distillation": 21411,
      "offering unified solution": 23302,
      "language models focusing": 16350,
      "models comprehensively understand": 21080,
      "models achieve stateoftheart": 20946,
      "effective prompt design": 8892,
      "reasoning tasks chainofthought": 27457,
      "hypothesize large language": 13968,
      "bert pretrained model": 3524,
      "accuracy privacy protection": 904,
      "models identify social": 21368,
      "language models extract": 16339,
      "zero fewshot performance": 35939,
      "zero fewshot settings": 35943,
      "language models evaluate": 16326,
      "llms llama models": 18771,
      "language models systematic": 16723,
      "points code available": 25072,
      "conduct set experiments": 5620,
      "modalities natural language": 20325,
      "language models reinforced": 16691,
      "existing opensource models": 10306,
      "llms substantial margin": 18976,
      "language models vlms": 16751,
      "extend large language": 10654,
      "models llms employed": 21518,
      "nlp tasks especially": 22952,
      "models similar scale": 21978,
      "interactions mental health": 15391,
      "information retrieval recommend": 14910,
      "software engineering tasks": 30456,
      "generated different models": 12353,
      "potential llms enhancing": 25275,
      "generate instruction data": 12292,
      "visionlanguage models lvlms": 35321,
      "settings zeroshot fewshot": 29747,
      "language models improves": 16384,
      "attains stateoftheart performance": 2701,
      "remains challenging existing": 27994,
      "context lengths gpt4": 5900,
      "finetuning prompt engineering": 11497,
      "visionlanguage models large": 35317,
      "models large visionlanguage": 21429,
      "specific domain knowledge": 30689,
      "question answering reasoning": 27052,
      "answering reasoning tasks": 2073,
      "experiments widely used": 10504,
      "approach achieves stateoftheart": 2229,
      "pretrained llms llama": 25707,
      "responses generated models": 28494,
      "code data public": 4736,
      "inference process involves": 14801,
      "enhances performance compared": 9552,
      "instruction tuning present": 15206,
      "current state art": 6531,
      "instructiontuned language models": 15285,
      "propose mechanism allows": 26527,
      "light pressing issue": 17832,
      "conducted experiments using": 5634,
      "model surpasses performance": 20818,
      "human preference data": 13854,
      "results evaluated gpt4": 28605,
      "shown remarkable capabilities": 29912,
      "different llms prompt": 8098,
      "llms gpt35 bard": 18650,
      "future research area": 12042,
      "downstream applications reducing": 8674,
      "language models producing": 16667,
      "represented training data": 28193,
      "expensive obtain paper": 10363,
      "paper address challenge": 24003,
      "approach yielded exceptional": 2362,
      "yielded exceptional results": 35916,
      "progress opensource large": 26224,
      "enhances model performance": 9548,
      "study highlights importance": 31339,
      "instruction tuning improve": 15193,
      "catastrophic forgetting multimodal": 4215,
      "forgetting multimodal large": 11734,
      "language models following": 16353,
      "catastrophic forgetting mllms": 4214,
      "image classification tasks": 14062,
      "llms billions parameters": 18450,
      "outstanding performance various": 23904,
      "model architecture design": 20375,
      "limited context window": 17947,
      "present novel framework": 25546,
      "sft training data": 29769,
      "sizes 7b 13b": 30294,
      "achieve average improvement": 944,
      "despite great success": 7781,
      "opensource llms llama2": 23524,
      "paper introduces new": 24063,
      "publicly available dataset": 26853,
      "training data opensource": 33485,
      "opensource code model": 23490,
      "models llms rely": 21615,
      "knowledge bases large": 15818,
      "bases large language": 3279,
      "approach mitigate challenges": 2315,
      "stateoftheart llms including": 30947,
      "llms including llama2": 18703,
      "demonstrate stateoftheart performance": 7498,
      "performance various multimodal": 24801,
      "including gpt2 gpt3": 14483,
      "humans work introduce": 13934,
      "generalpurpose large language": 12251,
      "llms paper introduce": 18839,
      "improve performance llms": 14285,
      "recent works demonstrated": 27572,
      "language model aligned": 16116,
      "offers effective efficient": 23308,
      "language modeling large": 16218,
      "address issue paper": 1333,
      "pretrained vision language": 25776,
      "context window size": 5928,
      "models llms requires": 21621,
      "natural language query": 22563,
      "data existing methods": 6700,
      "human effort required": 13803,
      "exploiting large language": 10560,
      "code llama code": 4773,
      "llms shown potential": 18942,
      "problems propose novel": 26033,
      "presents significant challenges": 25598,
      "research demonstrates effectiveness": 28303,
      "models mbert mt5": 21708,
      "matches outperforms existing": 19654,
      "process large language": 26070,
      "demonstrated remarkable proficiency": 7547,
      "commonly used benchmarks": 5030,
      "models realworld scenarios": 21897,
      "base model llama2": 3127,
      "pretrained models weights": 25729,
      "effectiveness wide applicability": 8976,
      "present new benchmark": 25542,
      "research paper introduces": 28340,
      "using small set": 34916,
      "language model developed": 16133,
      "opensourced code model": 23550,
      "generated gpt4 leads": 12363,
      "novel approach finetuning": 23059,
      "code pretrained models": 4799,
      "models llms garnered": 21536,
      "garnered considerable attention": 12121,
      "empirical results realworld": 9236,
      "training fewshot training": 33521,
      "models offer new": 21766,
      "code generation prompting": 4763,
      "code generated llms": 4753,
      "coding capabilities models": 4867,
      "generated candidates based": 12345,
      "poses challenging task": 25169,
      "designed automatically generate": 7725,
      "achieves strong performance": 1075,
      "models stable diffusion": 22008,
      "given input prompt": 12750,
      "harnessing large language": 13464,
      "language model palm": 16183,
      "generation dialogue systems": 12487,
      "models existing studies": 21231,
      "handle longer contexts": 13409,
      "models like clip": 21452,
      "tens thousands words": 32720,
      "touvron et al": 33305,
      "models llms help": 21554,
      "paper propose approach": 24107,
      "range tasks training": 27213,
      "finetuning opensource llms": 11467,
      "domains like medicine": 8629,
      "empirical study pretrained": 9244,
      "llms match surpass": 18796,
      "bilingual evaluation understudy": 3711,
      "reduces memory usage": 27737,
      "larger batch size": 17317,
      "techniques fall short": 32639,
      "a100 gpu hours": 617,
      "generation using large": 12632,
      "new framework called": 22806,
      "safe reinforcement learning": 29031,
      "presents significant challenge": 25597,
      "enhancing model performance": 9569,
      "commercial models chatgpt": 4991,
      "capabilities open source": 4056,
      "models llms despite": 21512,
      "visual recognition tasks": 35354,
      "does require access": 8537,
      "models recent work": 21906,
      "application large language": 2131,
      "performances various tasks": 24823,
      "provide public access": 26724,
      "knowledge distillation large": 15833,
      "generalization ability outofdistribution": 12207,
      "chatgpt large language": 4483,
      "establish strong baseline": 9771,
      "inject domain knowledge": 14980,
      "previous research shown": 25876,
      "code models data": 4788,
      "llms despite recent": 18533,
      "tasks results performance": 32490,
      "large vision language": 17288,
      "gpt llama families": 12854,
      "generation training procedure": 12624,
      "inspired recent success": 15100,
      "models llms natural": 21589,
      "lowresource languages bangla": 19314,
      "language tasks including": 16831,
      "training data observe": 33484,
      "al 2023 demonstrated": 1690,
      "learning icl ability": 17584,
      "current evaluation metrics": 6490,
      "news social media": 22886,
      "remarkable advancements recent": 28029,
      "advancements recent years": 1476,
      "humanlike text generation": 13910,
      "leading suboptimal performance": 17486,
      "finetuning results showcase": 11513,
      "wide range nlp": 35560,
      "significant improvements achieved": 29992,
      "using carefully curated": 34746,
      "execute complex instructions": 10192,
      "models llms combined": 21491,
      "llms specifically analyze": 18965,
      "applications various fields": 2183,
      "llama2 chatgpt gpt4": 18169,
      "speechbased slot filling": 30793,
      "various language tasks": 35108,
      "language tasks paper": 16832,
      "llms including gpt35turbo": 18696,
      "including gpt35turbo gpt4": 14488,
      "recent studies highlighted": 27555,
      "propose novel training": 26556,
      "novel training method": 23122,
      "models new data": 21751,
      "instruction tuning methods": 15203,
      "generate instructionfollowing data": 12294,
      "despite promising performance": 7804,
      "model llm pretraining": 20625,
      "models llms enhance": 21520,
      "llms knowledge distillation": 18735,
      "baselines zeroshot setting": 3276,
      "llms led widespread": 18753,
      "language models prone": 16669,
      "investigation large language": 15614,
      "models llms marked": 21583,
      "llms marked significant": 18794,
      "downstream tasks importantly": 8695,
      "text large language": 32904,
      "language model llmbased": 16172,
      "finetuning llama 7b": 11440,
      "finetune large language": 11286,
      "findings highlight potential": 11236,
      "models llms utilize": 21661,
      "programming languages python": 26202,
      "significantly improve accuracy": 30056,
      "model types llama": 20847,
      "ability handle longer": 683,
      "proprietary models like": 26650,
      "like chatgpt gpt4": 17854,
      "modern large language": 22162,
      "standard datasets models": 30874,
      "llms downstream applications": 18549,
      "llama2 7b 13b": 18163,
      "llms hidden states": 18670,
      "offers unique perspective": 23315,
      "enhance computational efficiency": 9510,
      "recognition ner tasks": 27639,
      "achieved best results": 1002,
      "hope work serve": 13759,
      "pursuit artificial general": 26890,
      "marked significant milestone": 19595,
      "language models rely": 16692,
      "training data work": 33494,
      "models using large": 22103,
      "prompting incontext learning": 26378,
      "strategy experimental results": 31121,
      "work present novel": 35749,
      "present novel method": 25547,
      "captioning visual question": 4139,
      "performance llms various": 24665,
      "experiments demonstrate proposed": 10434,
      "stable diffusion xl": 30850,
      "extensive experiments validate": 10704,
      "work introduce new": 35724,
      "average success rate": 3026,
      "tasks zeroshot prompting": 32557,
      "points exact match": 25074,
      "questionanswering tasks finetuning": 27090,
      "achieving highest accuracy": 1094,
      "promising direction enhancing": 26286,
      "llms gpt4 opensource": 18658,
      "gpt4 opensource counterparts": 13103,
      "necessary reproduce results": 22609,
      "results demonstrate model": 28590,
      "model natural language": 20653,
      "visual instruction datasets": 35339,
      "datasets generated large": 7122,
      "automatically generating natural": 2917,
      "frozen large language": 11937,
      "language model small": 16202,
      "small number parameters": 30363,
      "showcasing immense potential": 29851,
      "symbolic knowledge distillation": 31802,
      "models compared previous": 21073,
      "efficient effective method": 9032,
      "previous stateoftheart methods": 25882,
      "achieving significantly higher": 1102,
      "conducted benchmark datasets": 5630,
      "gpt 35 gpt": 12843,
      "methods paper presents": 20072,
      "key metric evaluating": 15777,
      "language models growing": 16376,
      "visual language model": 35342,
      "language model family": 16141,
      "knowledge language model": 15870,
      "moving average ema": 22241,
      "models incorporating external": 21387,
      "incorporating external knowledge": 14576,
      "commonly known hallucination": 5027,
      "relative position encoding": 27883,
      "pruning large language": 26811,
      "llms shown impressive": 18941,
      "state art model": 30905,
      "model codes available": 20427,
      "significant potential revolutionize": 30009,
      "techniques reinforcement learning": 32660,
      "conduct case study": 5586,
      "models zeroshot fewshot": 22144,
      "parameters achieve comparable": 24222,
      "llms demonstrate impressive": 18510,
      "impressive incontext learning": 14239,
      "current instruction tuning": 6494,
      "degrade model performance": 7380,
      "model performance address": 20696,
      "high computational memory": 13559,
      "achieves sota results": 1069,
      "marking significant advancement": 19602,
      "significant performance degradation": 30000,
      "utilizing external tools": 34971,
      "llms superior performance": 18983,
      "model llm output": 20624,
      "problem multimodal large": 25998,
      "current multimodal large": 6518,
      "code generation large": 4756,
      "language models codellms": 16279,
      "images experimental results": 14084,
      "proposed method outperforms": 26607,
      "models specifically designed": 22001,
      "language models perspective": 16651,
      "model aligned human": 20363,
      "human feedback extensive": 13819,
      "feedback extensive experiments": 11060,
      "experimental results chatgpt": 10390,
      "llms achieved stateoftheart": 18420,
      "learning multimodal large": 17616,
      "language model gpt35": 16152,
      "approach outperforms previous": 2322,
      "mathematical reasoning ability": 19683,
      "boosting large language": 3821,
      "empirical study large": 9241,
      "models llms significant": 21637,
      "llms significant advancements": 18948,
      "previous work focuses": 25893,
      "using lora method": 34829,
      "llms chatgpt demonstrated": 18464,
      "examples incontext learning": 10128,
      "models llms greatly": 21553,
      "incontext learning method": 14562,
      "ability llms generate": 699,
      "visionlanguage models like": 35320,
      "tasks face challenges": 32328,
      "bridge gap work": 3868,
      "recent development large": 27514,
      "smaller models flant5": 30389,
      "code data evaluation": 4730,
      "supervised finetuning models": 31675,
      "exhibits stateoftheart performance": 10253,
      "chainofthought prompting large": 4303,
      "models method requires": 21720,
      "models llms offers": 21595,
      "classification tasks using": 4617,
      "autoregressive language modeling": 2944,
      "gpt2 text generation": 12958,
      "unseen lowresource languages": 34442,
      "data lowresource languages": 6762,
      "et al 2023a": 9800,
      "instruction finetuning llms": 15159,
      "xu et al": 35885,
      "large language modelbased": 16988,
      "evidence support claim": 10065,
      "different llm architectures": 8095,
      "various linguistic tasks": 35112,
      "correlates human judgments": 6217,
      "exhibits comparable performance": 10245,
      "capability finetuned models": 4088,
      "graph language model": 13223,
      "visionlanguage model vlm": 35314,
      "approach does require": 2262,
      "terms bleu score": 32739,
      "future research application": 12041,
      "finetuning multimodal large": 11458,
      "instruction following data": 15164,
      "various model sizes": 35120,
      "model sizes notably": 20799,
      "annotated dataset available": 2019,
      "biases large language": 3676,
      "including zeroshot fewshot": 14531,
      "transforms natural language": 33807,
      "improve sample efficiency": 14299,
      "optimization large language": 23629,
      "language models gained": 16354,
      "models gained immense": 21301,
      "importance recent years": 14190,
      "questions remain unanswered": 27129,
      "context large language": 5896,
      "trainable parameters despite": 33381,
      "models llms dynamic": 21516,
      "responses generated llms": 28493,
      "rapid development large": 27248,
      "generalpurpose ai assistants": 12247,
      "diminishes attack success": 8204,
      "attack success rate": 2687,
      "success rate asr": 31524,
      "models llms solve": 21641,
      "llms tend generate": 19001,
      "using various prompt": 34938,
      "language model meets": 16174,
      "language models lvlms": 16613,
      "consists key steps": 5765,
      "model finetuning llama": 20535,
      "language models advanced": 16244,
      "models llms offer": 21594,
      "models paper propose": 21797,
      "results demonstrate superiority": 28596,
      "impressive capabilities multimodal": 14235,
      "evaluation pretrained models": 9993,
      "agents large language": 1571,
      "language models vs": 16754,
      "generate relevant accurate": 12318,
      "finally propose new": 11202,
      "llms increasingly used": 18713,
      "increasingly used various": 14648,
      "performs better using": 24846,
      "efforts align large": 9085,
      "propose novel technique": 26554,
      "novel technique called": 23117,
      "7b 13b 30b": 526,
      "incontext learning code": 14552,
      "code released github": 4807,
      "attributed key factors": 2761,
      "operations large language": 23572,
      "require access models": 28211,
      "models black box": 21031,
      "models llms improve": 21561,
      "potential using llms": 25306,
      "including chatgpt llama": 14466,
      "construct new evaluation": 5803,
      "using advanced large": 34728,
      "models llms generative": 21541,
      "better align human": 3588,
      "llms longer context": 18785,
      "longer context lengths": 19196,
      "comprehensive literature review": 5386,
      "language models verifiable": 16747,
      "despite significant progress": 7813,
      "study introduce novel": 31343,
      "key design decisions": 15761,
      "visionlanguage models multimodal": 35322,
      "models llms witnessed": 21665,
      "models finetuned llama": 21277,
      "problems varying difficulty": 26036,
      "mitigating hallucinations llms": 20264,
      "drug discovery process": 8749,
      "early results using": 8781,
      "significantly reduce cost": 30081,
      "lays solid foundation": 17459,
      "shown immense potential": 29886,
      "models llms especially": 21521,
      "conduct comprehensive evaluation": 5593,
      "13b parameter models": 131,
      "performance comparable chatgpt": 24548,
      "studies shown llms": 31285,
      "benchmarks demonstrate superiority": 3437,
      "multiple programming languages": 22411,
      "programming languages paper": 26201,
      "achieves comparable superior": 1041,
      "popular models like": 25130,
      "improving downstream accuracy": 14408,
      "llms wide range": 19045,
      "tasks deployment hindered": 32292,
      "significant performance drop": 30001,
      "scaling language models": 29166,
      "llms long context": 18783,
      "expensive training costs": 10369,
      "inputs 100k tokens": 15042,
      "achieves significant improvements": 1064,
      "recently demonstrated impressive": 27590,
      "commercial opensource llms": 4994,
      "models llms numerous": 21593,
      "leading opensource models": 17481,
      "commercial models gpt35": 4992,
      "gpt35 turbo gpt4": 13037,
      "zero fewshot prompts": 35941,
      "opensource models zeroshot": 23536,
      "shows opensource models": 29932,
      "data selection instruction": 6855,
      "selection instruction tuning": 29391,
      "acquiring highquality data": 1125,
      "introduce novel dataset": 15526,
      "approach inspired observation": 2301,
      "instruction tuning llama2": 15200,
      "superior performance general": 31652,
      "lack indepth understanding": 15994,
      "work highlights need": 35719,
      "various realworld scenarios": 35138,
      "including roberta gpt2": 14518,
      "extra inference cost": 10740,
      "evaluations multiple datasets": 10036,
      "languages training data": 16918,
      "gpt35 gpt4 generate": 13023,
      "multidocument question answering": 22278,
      "language models type": 16739,
      "large visual language": 17298,
      "models llms taken": 21647,
      "llms taken spotlight": 18991,
      "taken spotlight natural": 32027,
      "spotlight natural language": 30823,
      "language processing integrating": 16781,
      "processing integrating llms": 26103,
      "integrating llms vision": 15335,
      "llms vision enables": 19039,
      "vision enables users": 35296,
      "enables users explore": 9312,
      "users explore emergent": 34688,
      "explore emergent abilities": 10585,
      "models vlms llava": 22122,
      "vlms llava flamingo": 35379,
      "performance various visiolinguistic": 24809,
      "various visiolinguistic tasks": 35159,
      "visiolinguistic tasks consequently": 35288,
      "tasks consequently enormous": 32276,
      "consequently enormous applications": 5697,
      "enormous applications large": 9587,
      "applications large models": 2163,
      "large models potentially": 17239,
      "lack related work": 16000,
      "ability large models": 693,
      "extensive experiments showed": 10701,
      "achieves better overall": 1034,
      "opensource llms demonstrate": 23520,
      "current evaluation methods": 6489,
      "high quality synthetic": 13580,
      "larger models better": 17331,
      "way large language": 35441,
      "approach involves generating": 2303,
      "release model data": 27911,
      "performance multiple natural": 24685,
      "protein sequence generation": 26661,
      "models llms field": 21532,
      "advanced language models": 1424,
      "data benchmark comprises": 6623,
      "designed overcome challenges": 7739,
      "models llms process": 21607,
      "tasks including question": 32367,
      "including question answering": 14515,
      "advantages incontext learning": 1506,
      "language models users": 16744,
      "existing instruction tuning": 10277,
      "size training set": 30289,
      "study offers insights": 31371,
      "models llms remains": 21616,
      "address limitation introduce": 1341,
      "redteaming large language": 27696,
      "natural language applications": 22508,
      "generate test cases": 12330,
      "effective test cases": 8900,
      "compared existing methods": 5132,
      "stateoftheart vision transformers": 31007,
      "novel approach named": 23062,
      "emerging large language": 9194,
      "extensive experiments various": 10705,
      "models mllms recently": 21732,
      "gained immense popularity": 12063,
      "knowledge reasoning abilities": 15898,
      "models learn follow": 21440,
      "chinese large language": 4547,
      "existing llms llama": 10291,
      "scaling instruction tuning": 29164,
      "models despite remarkable": 21147,
      "novel efficient method": 23076,
      "input sequence length": 15028,
      "sheer number parameters": 29803,
      "parameters large language": 24262,
      "models llms open": 21596,
      "llms open new": 18827,
      "scenarios extensive experiments": 29205,
      "recent large visionlanguage": 27528,
      "reveal models demonstrate": 28805,
      "open foundation models": 23394,
      "extend context length": 10649,
      "model parameters using": 20692,
      "language models possess": 16657,
      "data significantly enhance": 6865,
      "scarcity publicly available": 29196,
      "instructiontuned llama models": 15293,
      "achieve significant performance": 983,
      "low computational overhead": 19267,
      "models ability capture": 20933,
      "stateoftheart competitive performance": 30927,
      "offers valuable insights": 23318,
      "performance generalpurpose llms": 24615,
      "finetuned opensource llms": 11349,
      "like gpt4 llama": 17874,
      "models perform data": 21809,
      "attention patterns early": 2732,
      "patterns early layers": 24410,
      "high training costs": 13587,
      "performance paper propose": 24710,
      "increase computational overhead": 14594,
      "deployment large language": 7652,
      "llama2 mistral models": 18185,
      "align llms human": 1733,
      "data selection method": 6857,
      "et al 2023b": 9801,
      "algorithms language models": 1720,
      "simple effective strategy": 30146,
      "models consistently outperform": 21101,
      "stateoftheart pretrained models": 30978,
      "llms demonstrated great": 18514,
      "demonstrated great potential": 7524,
      "models llms presents": 21605,
      "llms garnered significant": 18629,
      "llama alpaca vicuna": 18074,
      "demonstrates strong zeroshot": 7574,
      "remains significant gap": 28014,
      "models llms able": 21474,
      "visual instruction data": 35338,
      "27 billion parameters": 283,
      "opening new avenues": 23468,
      "indicate models currently": 14691,
      "model better understand": 20400,
      "rtx 2080 ti": 29001,
      "llms findings indicate": 18609,
      "presents set challenges": 25595,
      "capabilities llms significant": 4041,
      "performance compared models": 24551,
      "criteria experimental results": 6378,
      "effective method enhance": 8884,
      "cover diverse set": 6317,
      "natural language description": 22513,
      "study demonstrate potential": 31315,
      "designing data methods": 7753,
      "data methods effective": 6768,
      "opensourced large language": 23553,
      "applications code models": 2145,
      "significantly outperform larger": 30071,
      "attention mechanism transformer": 2726,
      "computational overhead work": 5474,
      "models extensive experiments": 21248,
      "extensive experiments diverse": 10698,
      "way future research": 35434,
      "enhance reasoning capabilities": 9530,
      "response challenge present": 28473,
      "conduct indepth analysis": 5613,
      "parameters finetuning large": 24247,
      "model efficient inference": 20483,
      "hope proposed method": 13756,
      "computational costs associated": 5462,
      "number input tokens": 23146,
      "pretraining supervised finetuning": 25843,
      "models llms opened": 21597,
      "llms opened new": 18831,
      "findings reveal llms": 11249,
      "training data long": 33480,
      "degree language models": 7388,
      "gpt4 opensource models": 13104,
      "nlp transformerbased models": 22967,
      "compared widely used": 5183,
      "language models tackle": 16726,
      "performance widely used": 24817,
      "conducted extensive empirical study": 5637,
      "pretrained language models existing": 25670,
      "language model gpt2 generate": 16150,
      "performs better par stateoftheart": 24845,
      "generation using pretrained language": 12636,
      "pretrained language models shown": 25693,
      "language models shown remarkable": 16704,
      "text pretrained language models": 32919,
      "language models largescale language": 16411,
      "models largescale language models": 21434,
      "large language models pretrained": 17187,
      "pretrained language models finetuning": 25672,
      "text generation model gpt2": 32877,
      "paper presents novel approach": 24104,
      "pretrained language models capable": 25667,
      "language models capable generating": 16271,
      "work propose new method": 35755,
      "large language models shown": 17201,
      "results showed finetuned model": 28680,
      "representations transformers bert generative": 28177,
      "language processing nlp recently": 16793,
      "transformerbased models bert gpt2": 33762,
      "finetuning large pretrained language": 11434,
      "downstream natural language processing": 8680,
      "experimental results proposed method": 10406,
      "language models gpt3 t5": 16371,
      "pretrained language model t5": 25662,
      "models like bert gpt2": 21449,
      "tasks finetuning pretrained models": 32340,
      "achieves comparable better performance": 1038,
      "recent work shown large": 27570,
      "work shown large language": 35783,
      "shown large language models": 29895,
      "providing natural language instructions": 26778,
      "largescale language models llms": 17361,
      "test set best model": 32786,
      "language models plms gpt2": 16655,
      "using natural language processing": 34852,
      "ability pretrained language models": 714,
      "language generation need training": 16088,
      "stateoftheart natural language generation": 30964,
      "natural language generation models": 22522,
      "area natural language processing": 2480,
      "large number trainable parameters": 17256,
      "paper propose simple effective": 24117,
      "train large language model": 33366,
      "models propose new paradigm": 21867,
      "zeroshot capabilities large language": 35956,
      "large language models identify": 17044,
      "large language model families": 16966,
      "received little attention paper": 27480,
      "baseline future research code": 3247,
      "prompt tuning prompt tuning": 26354,
      "demonstrated superior performance various": 7553,
      "language models exhibit emergent": 16331,
      "arithmetic commonsense symbolic reasoning": 2502,
      "transformers large language models": 33790,
      "large language models like": 17052,
      "language models like gpt35": 16418,
      "language models plms t5": 16656,
      "scale large language models": 29139,
      "work shown finetuning large": 35781,
      "paper propose novel approach": 24114,
      "largescale pretrained language model": 17375,
      "language models increasingly popular": 16390,
      "outperform larger language models": 23778,
      "similar large language models": 30108,
      "models natural language processing": 21746,
      "pretrained natural language models": 25737,
      "larger language models trained": 17324,
      "visual question answering vqa": 35351,
      "language models lms increasingly": 16602,
      "language models llms gpt35": 16487,
      "large multilingual language model": 17243,
      "approach outperforms stateoftheart methods": 2325,
      "language models llms popular": 16536,
      "large language models generate": 17032,
      "recent proliferation large language": 27547,
      "model works phases phase": 20874,
      "proximal policy optimization algorithm": 26796,
      "efficient finetuning language models": 9036,
      "approximation fisher information matrix": 2421,
      "models llms including chatgpt": 21563,
      "large language models perform": 17184,
      "large language models text": 17209,
      "language models text generation": 16731,
      "instruction tuning instruction tuning": 15195,
      "language models llms using": 16587,
      "models llms using machinegenerated": 21659,
      "llms using machinegenerated instructionfollowing": 19028,
      "using machinegenerated instructionfollowing data": 34835,
      "zeroshot capabilities new tasks": 35959,
      "paper present attempt use": 24087,
      "revolutionizing natural language processing": 28856,
      "language models llms showcased": 16559,
      "models llms showcased remarkable": 21628,
      "remarkable performance diverse domains": 28046,
      "output large language models": 23871,
      "language models llms human": 16496,
      "challenges propose novel approach": 4373,
      "models llms demonstrated significant": 21508,
      "encoder visionlanguage models vlms": 9360,
      "language models llms focusing": 16475,
      "open pretrained transformers opt": 23414,
      "open large language model": 23402,
      "language models finetuning pretrained": 16348,
      "models llms chatgpt openai": 21490,
      "paper make attempt investigate": 24079,
      "models llms human values": 21559,
      "large language models detecting": 17011,
      "language models llms explore": 16471,
      "incontext learning incontext learning": 14559,
      "alpaca experimental results demonstrate": 1830,
      "models llms significantly enhanced": 21640,
      "study provides valuable insights": 31387,
      "decoderonly large language models": 7259,
      "improve language model efficiency": 14272,
      "modern pretrained language models": 22171,
      "widespread use language models": 35599,
      "results human evaluation demonstrate": 28625,
      "demonstrate effectiveness proposed method": 7450,
      "building better base models": 3921,
      "information large language models": 14879,
      "augmenting large language models": 2827,
      "using lowrank adaptation lora": 34832,
      "language model llm finetuned": 16167,
      "language models lms typically": 16610,
      "powerful large language models": 25344,
      "models llms gpt llama2": 21544,
      "language models llms difficult": 16455,
      "truthfulness large language models": 33942,
      "language models demonstrated ability": 16300,
      "like large language models": 17881,
      "language models llms work": 16594,
      "models foundation models fms": 21295,
      "investigate ability pretrained language": 15573,
      "abstract meaning representation amr": 772,
      "pretrained transformer gpt models": 25758,
      "domain recent advancements language": 8589,
      "language models lms led": 16603,
      "exceptional capabilities wide range": 10169,
      "language models llms based": 16433,
      "align large language models": 1730,
      "parameterefficient finetuning peft techniques": 24211,
      "integration large language models": 15342,
      "large language model large": 16969,
      "adopt curriculum learning strategy": 1403,
      "llms demonstrated remarkable abilities": 18521,
      "biomedical natural language processing": 3747,
      "models zero fewshot scenarios": 22142,
      "generative visionlanguage models vlms": 12715,
      "hallucinations large language models": 13393,
      "state art natural language": 30907,
      "art natural language processing": 2518,
      "performance various downstream tasks": 24800,
      "language reasoning capabilities large": 16814,
      "language models llm foundation": 16423,
      "models llm foundation models": 21471,
      "distillation large language models": 8342,
      "capabilities large language model": 4033,
      "results demonstrate effectiveness proposed": 28589,
      "wide range downstream tasks": 35553,
      "hypothesize large language models": 13969,
      "language models identify social": 16381,
      "large language models evaluate": 17017,
      "large language models systematic": 17207,
      "extend large language models": 10655,
      "language models llms employed": 16460,
      "large language models knowledge": 17047,
      "large visionlanguage models large": 17293,
      "visionlanguage models large visionlanguage": 35318,
      "models large visionlanguage models": 21430,
      "large visionlanguage models lvlms": 17295,
      "question answering reasoning tasks": 27053,
      "llms shown remarkable capabilities": 18945,
      "approach yielded exceptional results": 2363,
      "catastrophic forgetting multimodal large": 4216,
      "forgetting multimodal large language": 11735,
      "large language models following": 17028,
      "demonstrated outstanding performance various": 7538,
      "language models like llama": 16419,
      "using publicly available dataset": 34892,
      "opensource code model data": 23491,
      "language models llms rely": 16549,
      "external knowledge bases large": 10730,
      "knowledge bases large language": 15819,
      "bases large language models": 3280,
      "demonstrate stateoftheart performance various": 7499,
      "generalpurpose large language models": 12252,
      "llms achieved remarkable performance": 18418,
      "models recent works demonstrated": 21908,
      "large language model aligned": 16961,
      "shown remarkable capabilities various": 29913,
      "data experimental results demonstrate": 6704,
      "language models llms requires": 16554,
      "exploiting large language models": 10561,
      "models llms shown potential": 21633,
      "process large language models": 26071,
      "pretrained language models instruction": 25675,
      "opensourced code model weights": 23551,
      "language models llms garnered": 16478,
      "language models offer new": 16638,
      "paper present novel approach": 24092,
      "harnessing large language models": 13465,
      "demonstrated remarkable performance various": 7545,
      "language models existing studies": 16333,
      "touvron et al 2023": 33306,
      "language models llms help": 16494,
      "integrating large language models": 15333,
      "generation using large language": 12633,
      "language models llms training": 16581,
      "safe reinforcement learning human": 29032,
      "language models llms despite": 16454,
      "language models recent work": 16687,
      "application large language models": 2132,
      "knowledge distillation large language": 15834,
      "chatgpt large language models": 4484,
      "experimental results demonstrate proposed": 10396,
      "experimental results demonstrate approach": 10392,
      "large vision language models": 17289,
      "wang et al 2022": 35417,
      "et al 2023 demonstrated": 9799,
      "incontext learning icl ability": 14556,
      "remarkable advancements recent years": 28030,
      "wide range nlp tasks": 35561,
      "language models llms combined": 16440,
      "llms including gpt35turbo gpt4": 18697,
      "propose novel training method": 26557,
      "language model llm pretraining": 16170,
      "llms demonstrated superior performance": 18528,
      "language models llms enhance": 16462,
      "models llms led widespread": 21573,
      "language models llms marked": 16520,
      "models llms marked significant": 21584,
      "large language model llmbased": 16979,
      "finetune large language models": 11287,
      "language models llms utilize": 16589,
      "modern large language models": 22163,
      "entity recognition ner tasks": 9650,
      "pursuit artificial general intelligence": 26891,
      "finetuning large pretrained models": 11436,
      "large language models propose": 17189,
      "image captioning visual question": 14059,
      "captioning visual question answering": 4140,
      "extensive experiments demonstrate proposed": 10696,
      "work propose novel approach": 35757,
      "llms gpt4 opensource counterparts": 18659,
      "paper propose new benchmark": 24112,
      "experimental results demonstrate model": 10395,
      "models llms focusing llama": 21534,
      "datasets generated large language": 7123,
      "automatically generating natural language": 2918,
      "large language model small": 16986,
      "outperforms previous stateoftheart methods": 23845,
      "large language model specifically": 16987,
      "visual language models visual": 35344,
      "visual language models vlms": 35345,
      "large language models growing": 17043,
      "models incorporating external knowledge": 21388,
      "pruning large language models": 26812,
      "models llms shown impressive": 21632,
      "techniques reinforcement learning human": 32661,
      "models zeroshot fewshot settings": 22145,
      "parameters achieve comparable performance": 24223,
      "language model llm output": 16169,
      "problem multimodal large language": 25999,
      "current multimodal large language": 6519,
      "code generation large language": 4757,
      "large language models codellms": 17000,
      "proposed method outperforms stateoftheart": 26608,
      "models llms including gpt35": 21564,
      "learning human feedback extensive": 17577,
      "human feedback extensive experiments": 13820,
      "learning multimodal large language": 17617,
      "large language model gpt35": 16967,
      "approach outperforms previous stateoftheart": 2323,
      "boosting large language model": 3822,
      "empirical study large language": 9242,
      "language models llms significant": 16567,
      "models llms significant advancements": 21638,
      "language models llms greatly": 16493,
      "language models llms offers": 16530,
      "xu et al 2023": 35886,
      "graph language model glm": 13224,
      "finetuning multimodal large language": 11459,
      "biases large language models": 3677,
      "optimization large language models": 23630,
      "large language models gained": 17029,
      "language models gained immense": 16355,
      "context large language models": 5897,
      "language models llms dynamic": 16458,
      "rapid development large language": 27249,
      "diminishes attack success rate": 8205,
      "attack success rate asr": 2688,
      "language models llms solve": 16571,
      "vision language models lvlms": 35304,
      "extensive experimental results demonstrate": 10691,
      "large language models advanced": 16992,
      "language models llms offer": 16529,
      "stateoftheart llms including gpt4": 30948,
      "large language models vs": 17219,
      "large language models demonstrated": 17007,
      "models llms increasingly used": 21568,
      "efforts align large language": 9086,
      "propose novel technique called": 26555,
      "agents large language models": 1572,
      "operations large language models": 23573,
      "language models llms improve": 16499,
      "llms including chatgpt llama": 18691,
      "using advanced large language": 34729,
      "language models llms generative": 16482,
      "large language models verifiable": 17218,
      "large visionlanguage models multimodal": 17296,
      "language models llms witnessed": 16593,
      "language models llms especially": 16463,
      "recent studies shown llms": 27557,
      "including gpt35turbo gpt4 llama2": 14489,
      "language models llms numerous": 16528,
      "data selection instruction tuning": 6856,
      "models llms achieved stateoftheart": 21481,
      "llms achieved stateoftheart performance": 18421,
      "large language models type": 17214,
      "large visual language models": 17299,
      "language models llms taken": 16577,
      "models llms taken spotlight": 21648,
      "llms taken spotlight natural": 18992,
      "taken spotlight natural language": 32028,
      "spotlight natural language processing": 30824,
      "natural language processing integrating": 22541,
      "language processing integrating llms": 16782,
      "processing integrating llms vision": 26104,
      "integrating llms vision enables": 15336,
      "llms vision enables users": 19040,
      "vision enables users explore": 35297,
      "enables users explore emergent": 9313,
      "users explore emergent abilities": 34689,
      "language models vlms llava": 16752,
      "models vlms llava flamingo": 22123,
      "impressive performance various visiolinguistic": 14244,
      "performance various visiolinguistic tasks": 24810,
      "various visiolinguistic tasks consequently": 35160,
      "visiolinguistic tasks consequently enormous": 35289,
      "tasks consequently enormous applications": 32277,
      "consequently enormous applications large": 5698,
      "enormous applications large models": 9588,
      "applications large models potentially": 2164,
      "large models potentially used": 17240,
      "way large language models": 35442,
      "performance multiple natural language": 24686,
      "language models llms field": 16474,
      "language models llms process": 16542,
      "tasks including question answering": 32368,
      "existing instruction tuning datasets": 10278,
      "models llms like gpt3": 21577,
      "language models llms remains": 16550,
      "models llms remains significant": 21617,
      "llms remains significant challenge": 18907,
      "redteaming large language models": 27697,
      "propose novel approach named": 26546,
      "emerging large language models": 9195,
      "language models mllms recently": 16625,
      "opensource llms llama2 mistral": 23525,
      "language models despite remarkable": 16306,
      "parameters large language models": 24263,
      "language models llms open": 16531,
      "recent large visionlanguage models": 27529,
      "large models like gpt4": 17237,
      "attention patterns early layers": 2733,
      "model performance paper propose": 20705,
      "impressive incontext learning icl": 14240,
      "deployment large language models": 7653,
      "models llms demonstrated great": 21499,
      "llms demonstrated great potential": 18515,
      "language models llms presents": 16540,
      "models llms garnered significant": 21537,
      "llms garnered significant attention": 18630,
      "language models llms able": 16427,
      "emerged effective method enhance": 9159,
      "designing data methods effective": 7754,
      "applications code models available": 2146,
      "parameters finetuning large language": 24248,
      "language models llms opened": 16532,
      "models llms opened new": 21598,
      "large language models produce": 17188,
      "large pretrained language models bert": 17262,
      "generation using pretrained language models": 12637,
      "language models largescale language models": 16412,
      "large pretrained language models capable": 17263,
      "advances natural language processing nlp": 1486,
      "encoder representations transformers bert generative": 9357,
      "natural language processing nlp recently": 22551,
      "finetuning large pretrained language models": 11435,
      "downstream natural language processing nlp": 8681,
      "work shown large language models": 35784,
      "pretrained language models plms gpt2": 25687,
      "zeroshot capabilities large language models": 35957,
      "shown large language models llms": 29896,
      "large language models like gpt35": 17053,
      "pretrained language models plms t5": 25688,
      "models natural language processing nlp": 21747,
      "large language models llms gpt35": 17095,
      "large language models llms popular": 17127,
      "recent proliferation large language models": 27548,
      "language models llms including chatgpt": 16501,
      "large language models text generation": 17210,
      "large language models llms using": 17163,
      "language models llms using machinegenerated": 16588,
      "models llms using machinegenerated instructionfollowing": 21660,
      "llms using machinegenerated instructionfollowing data": 19029,
      "performance natural language processing nlp": 24690,
      "tasks named entity recognition ner": 32426,
      "large language models llms showcased": 17144,
      "language models llms showcased remarkable": 16560,
      "output large language models llms": 23872,
      "large language models llms human": 17101,
      "language models llms demonstrated significant": 16451,
      "large language models llms focusing": 17086,
      "language models llms chatgpt openai": 16439,
      "aligning large language models llms": 1748,
      "language models llms human values": 16497,
      "language models llms significantly enhanced": 16570,
      "benchmarking large language models fewshot": 3426,
      "information large language models llms": 14880,
      "large language model llm finetuned": 16974,
      "powerful large language models llms": 25345,
      "language models llms gpt llama2": 16485,
      "large language models llms difficult": 17071,
      "large language models llms work": 17168,
      "investigate ability pretrained language models": 15574,
      "generative pretrained transformer gpt models": 12694,
      "large language models llms based": 17061,
      "recent large language models llm": 27526,
      "integration large language models llms": 15343,
      "large language model large language": 16970,
      "finetuned large language models llms": 11328,
      "state art natural language processing": 30908,
      "language reasoning capabilities large language": 16815,
      "large language models llm foundation": 17056,
      "language models llm foundation models": 16424,
      "experimental results demonstrate effectiveness proposed": 10394,
      "prompting large language models llms": 26385,
      "large language models llms employed": 17076,
      "performance large language models generate": 24647,
      "large visionlanguage models large visionlanguage": 17294,
      "visionlanguage models large visionlanguage models": 35319,
      "models llms shown remarkable capabilities": 21636,
      "remarkable capabilities natural language processing": 28035,
      "catastrophic forgetting multimodal large language": 4217,
      "forgetting multimodal large language models": 11736,
      "transformerbased large language models llms": 33756,
      "large language models llms rely": 17137,
      "external knowledge bases large language": 10731,
      "knowledge bases large language models": 15820,
      "bases large language models llms": 3281,
      "generalpurpose large language models llms": 12253,
      "models llms achieved remarkable performance": 21479,
      "size large language models llms": 30258,
      "large language models llms requires": 17141,
      "language models llms shown potential": 16564,
      "large language models llms garnered": 17088,
      "pretrained language models existing studies": 25671,
      "large language models llms help": 17099,
      "era large language models llms": 9700,
      "large language models llms training": 17157,
      "safe reinforcement learning human feedback": 29033,
      "large language models llms despite": 17070,
      "application large language models llms": 2133,
      "knowledge distillation large language models": 15835,
      "chatgpt large language models llms": 4485,
      "large language models llms combined": 17065,
      "large language model llm pretraining": 16977,
      "models llms demonstrated superior performance": 21511,
      "large language models llms enhance": 17078,
      "language models llms led widespread": 16511,
      "large language models llms marked": 17114,
      "language models llms marked significant": 16521,
      "large language models llms utilize": 17164,
      "named entity recognition ner tasks": 22488,
      "image captioning visual question answering": 14060,
      "language models llms focusing llama": 16476,
      "datasets generated large language models": 7124,
      "language models llms shown impressive": 16563,
      "techniques reinforcement learning human feedback": 32662,
      "large language model llm output": 16976,
      "current multimodal large language models": 6520,
      "code generation large language models": 4758,
      "multimodal large language models llms": 22357,
      "language models llms including gpt35": 16502,
      "align large language models llms": 1731,
      "reinforcement learning human feedback extensive": 27843,
      "learning human feedback extensive experiments": 17578,
      "empirical study large language models": 9243,
      "large language models llms significant": 17146,
      "language models llms significant advancements": 16568,
      "large language models llms greatly": 17098,
      "large language models llms offers": 17122,
      "finetuning multimodal large language models": 11460,
      "large language models gained immense": 17030,
      "large language models llms dynamic": 17074,
      "rapid development large language models": 27250,
      "diminishes attack success rate asr": 8206,
      "large language models llms solve": 17148,
      "large vision language models lvlms": 17290,
      "large language models llms offer": 17121,
      "advent large language models llms": 1512,
      "language models llms increasingly used": 16506,
      "efforts align large language models": 9087,
      "large language models llms improve": 17103,
      "using advanced large language models": 34730,
      "large language models llms generative": 17091,
      "large language models llms witnessed": 17167,
      "finetuning large language model llm": 11430,
      "large language models llms especially": 17079,
      "llms including gpt35turbo gpt4 llama2": 18698,
      "large language models llms numerous": 17120,
      "language models llms achieved stateoftheart": 16431,
      "models llms achieved stateoftheart performance": 21482,
      "large language models llms taken": 17154,
      "language models llms taken spotlight": 16578,
      "models llms taken spotlight natural": 21649,
      "llms taken spotlight natural language": 18993,
      "taken spotlight natural language processing": 32029,
      "spotlight natural language processing integrating": 30825,
      "natural language processing integrating llms": 22542,
      "language processing integrating llms vision": 16783,
      "processing integrating llms vision enables": 26105,
      "integrating llms vision enables users": 15337,
      "llms vision enables users explore": 19041,
      "vision enables users explore emergent": 35298,
      "enables users explore emergent abilities": 9314,
      "visual language models vlms llava": 35346,
      "language models vlms llava flamingo": 16753,
      "demonstrated impressive performance various visiolinguistic": 7531,
      "impressive performance various visiolinguistic tasks": 14245,
      "performance various visiolinguistic tasks consequently": 24811,
      "various visiolinguistic tasks consequently enormous": 35161,
      "visiolinguistic tasks consequently enormous applications": 35290,
      "tasks consequently enormous applications large": 32278,
      "consequently enormous applications large models": 5699,
      "enormous applications large models potentially": 9589,
      "applications large models potentially used": 2165,
      "large language models llms field": 17085,
      "large language models llms process": 17132,
      "language models llms like gpt3": 16514,
      "large language models llms remains": 17138,
      "language models llms remains significant": 16551,
      "models llms remains significant challenge": 21618,
      "emerging large language models llms": 9196,
      "large language models mllms recently": 17174,
      "large language models despite remarkable": 17010,
      "parameters large language models llms": 24264,
      "large language models llms open": 17123,
      "language models llms demonstrated great": 16448,
      "models llms demonstrated great potential": 21500,
      "large language models llms presents": 17130,
      "language models llms garnered significant": 16479,
      "models llms garnered significant attention": 21538,
      "large language models llms able": 17058,
      "large language models paper presents": 17182,
      "parameters finetuning large language models": 24249,
      "large language models llms opened": 17124,
      "language models llms opened new": 16533,
      "vulnerabilities": 35403,
      "gem": 12137,
      "inherited": 14955,
      "ide": 13983,
      "top5": 33279,
      "editors": 8832,
      "statically": 31015,
      "vulnerable": 35405,
      "cycles": 6575,
      "regulation": 27831,
      "adversary": 1522,
      "countermeasures": 6300,
      "white": 35540,
      "similarlysized": 30136,
      "threat": 33089,
      "adversaries": 1521,
      "defense": 7359,
      "triggered": 33899,
      "interoperability": 15450,
      "managers": 19530,
      "strike": 31150,
      "datadependent": 6924,
      "regenerate": 27814,
      "blocksparse": 3779,
      "dangerous": 6583,
      "auditing": 2786,
      "eyetracking": 10812,
      "321": 320,
      "signatures": 29946,
      "radar": 27148,
      "groupwise": 13307,
      "clipping": 4668,
      "clipped": 4667,
      "wall": 35410,
      "worstcase": 35841,
      "nlcode": 22906,
      "12b": 104,
      "nlms": 22917,
      "spam": 30588,
      "nlm": 22916,
      "outlining": 23745,
      "pii": 24951,
      "javascript": 15702,
      "watermarking": 35425,
      "invisible": 15620,
      "selfcontained": 29414,
      "tricks": 33895,
      "codebases": 4835,
      "median": 19763,
      "monetary": 22206,
      "layouts": 17455,
      "ungrammatical": 34315,
      "contextrelated": 5933,
      "programmers": 26197,
      "tensorflow": 32723,
      "alpacas": 1840,
      "duplicated": 8752,
      "prefixlm": 25484,
      "say": 29119,
      "subtly": 31499,
      "violation": 35280,
      "tracing": 33325,
      "subroutines": 31444,
      "gpt2like": 12974,
      "13m": 133,
      "16gb": 166,
      "intend": 15366,
      "se": 29299,
      "stereotypes": 31061,
      "evade": 9813,
      "incentivized": 14442,
      "bullet": 3940,
      "workinprogress": 35806,
      "analysts": 1981,
      "unmodified": 34415,
      "layernorm": 17435,
      "attached": 2683,
      "leq": 17719,
      "preexisting": 25464,
      "lowrisk": 19319,
      "union": 34353,
      "462": 399,
      "started": 30898,
      "press": 25614,
      "orchestrating": 23666,
      "positives": 25199,
      "lightgbm": 17835,
      "mislabeled": 20215,
      "cyberphysical": 6573,
      "cuda": 6454,
      "contracts": 6004,
      "contract": 6003,
      "behaviours": 3333,
      "refuse": 27805,
      "autogenerated": 2848,
      "capitalizing": 4132,
      "located": 19145,
      "piqa": 24973,
      "unaffected": 34103,
      "shadow": 29772,
      "compromised": 5437,
      "untrustworthy": 34463,
      "consumes": 5821,
      "enterprises": 9620,
      "scattered": 29197,
      "protects": 26658,
      "protecting": 26656,
      "linux": 18032,
      "parrots": 24314,
      "victim": 35245,
      "harbor": 13419,
      "privacysensitive": 25956,
      "server": 29652,
      "optimistic": 23620,
      "grant": 13213,
      "misused": 20245,
      "censorship": 4262,
      "refusal": 27804,
      "degradations": 7377,
      "425": 387,
      "representatives": 28189,
      "provoke": 26789,
      "gigabytes": 12735,
      "perceivable": 24454,
      "indistinguishable": 14712,
      "affine": 1545,
      "jigsaw": 15706,
      "propensity": 26467,
      "pinpointed": 24957,
      "282": 287,
      "chatgptgpt4": 4505,
      "codegenmono16b": 4840,
      "freezes": 11919,
      "conservative": 5700,
      "llama34b": 18226,
      "persuasive": 24903,
      "tdd": 32577,
      "145": 139,
      "postpruning": 25229,
      "narrower": 22494,
      "apple": 2116,
      "llamacpp": 18239,
      "container": 5832,
      "weakened": 35457,
      "deceiving": 7223,
      "uploaded": 34480,
      "fortifies": 11779,
      "collectively": 4937,
      "intricately": 15486,
      "likelihoodbased": 17900,
      "mplugowl2": 22244,
      "encouraged": 9399,
      "contributors": 6046,
      "investment": 15618,
      "violence": 35282,
      "suicide": 31592,
      "author": 2834,
      "timing": 33170,
      "authorship": 2838,
      "magnitudes": 19385,
      "optimizationbased": 23638,
      "inserts": 15061,
      "forest": 11728,
      "scalings": 29182,
      "doubles": 8668,
      "reevaluating": 27769,
      "156": 148,
      "focal": 11642,
      "harming": 13445,
      "a6000": 621,
      "48gb": 407,
      "unraveling": 34423,
      "presentation": 25564,
      "branches": 3854,
      "spill": 30808,
      "kaggle": 15738,
      "landscapes": 16028,
      "defines": 7365,
      "textdavinci": 32978,
      "floatingpoint": 11623,
      "caching": 3960,
      "payload": 24426,
      "connectivity": 5687,
      "accesses": 830,
      "equivariance": 9689,
      "permuted": 24870,
      "dream": 8732,
      "resnet50": 28399,
      "hessian": 13528,
      "scanning": 29184,
      "trainers": 33436,
      "stringent": 31155,
      "groupedquery": 13299,
      "rotary": 28973,
      "projectspecific": 26256,
      "sequences tokens": 29616,
      "optimized using": 23648,
      "code completion": 4726,
      "largescale deep": 17348,
      "model discuss": 20471,
      "top5 accuracy": 33280,
      "files model": 11164,
      "years witnessed": 35900,
      "threats posed": 33092,
      "highly relevant": 13667,
      "image based": 14056,
      "transformers gpt2": 33783,
      "architectures datasets": 2461,
      "methods natural": 20068,
      "26 million": 279,
      "datasets paper": 7159,
      "identifiable information": 13994,
      "generating fake": 12422,
      "generate fake": 12277,
      "transparency ai": 33862,
      "potential harm": 25258,
      "algorithm gpt2": 1707,
      "software evolution": 30457,
      "source information": 30560,
      "vector embeddings": 35194,
      "representations bert": 28157,
      "opens possibilities": 23480,
      "network dnn": 22688,
      "dnn models": 8497,
      "models lower": 21691,
      "model utility": 20854,
      "modern machine": 22166,
      "code solutions": 4817,
      "difficult accurately": 8169,
      "single token": 30225,
      "reduce overall": 27723,
      "second design": 29320,
      "automatically identify": 2920,
      "closer real": 4697,
      "including difficulty": 14472,
      "faster algorithms": 10996,
      "memory cost": 19810,
      "maintain accuracy": 19414,
      "platforms using": 25012,
      "gpt2 generative": 12898,
      "posts using": 25231,
      "gpt natural": 12858,
      "strike balance": 31151,
      "words neural": 35659,
      "allows control": 1810,
      "t5 experiments": 31943,
      "based type": 3234,
      "models memorize": 21714,
      "nexttoken predictions": 22895,
      "tools detect": 33269,
      "plagiarism detection": 24988,
      "detection tool": 7885,
      "detection techniques": 7884,
      "requirements paper": 28241,
      "samples language": 29078,
      "opensource existing": 23501,
      "achieve close": 951,
      "parameters based": 24227,
      "results small": 28685,
      "models advances": 20968,
      "advances stateoftheart": 1487,
      "open access": 23382,
      "source training": 30571,
      "prompts analysis": 26404,
      "make training": 19486,
      "investigate approach": 15576,
      "issues using": 15675,
      "especially effective": 9733,
      "tends improve": 32714,
      "quality reduce": 26975,
      "compression recent": 5427,
      "llms bert": 18447,
      "deployed specific": 7637,
      "compression propose": 5423,
      "sparsity levels": 30633,
      "gptneo gptj": 13157,
      "exceeds performance": 10152,
      "trivial task": 33920,
      "effective mitigating": 8886,
      "online safety": 23369,
      "processing applications": 26094,
      "applications use": 2180,
      "data access": 6586,
      "extracting meaningful": 10758,
      "weights used": 35517,
      "work compares": 35678,
      "processing approaches": 26095,
      "code including": 4767,
      "traditionally used": 33356,
      "models ensure": 21208,
      "module integrate": 22193,
      "similar original": 30112,
      "provide affirmative": 26683,
      "time overhead": 33137,
      "learning memoryefficient": 17601,
      "underperform standard": 34166,
      "training epoch": 33510,
      "wall time": 35411,
      "175 billionparameter": 174,
      "gpt2 summarization": 12955,
      "designed natural": 7736,
      "semantic meaning": 29460,
      "meaning original": 19720,
      "completion tasks": 5263,
      "better robustness": 3627,
      "generation extend": 12502,
      "nlcode pairs": 22907,
      "models substantial": 22021,
      "fewer errors": 11087,
      "models nlms": 21753,
      "difficult detect": 8171,
      "information pii": 14897,
      "11b parameter": 85,
      "evaluate multiple": 9853,
      "potential harms": 25259,
      "negligible impact": 22671,
      "using efficient": 34771,
      "model api": 20368,
      "model open": 20662,
      "evaluating models": 9907,
      "benchmark named": 3401,
      "generating code": 12413,
      "models advance": 20966,
      "improving training": 14425,
      "candidates potential": 3996,
      "success training": 31527,
      "weights training": 35515,
      "achieving strong": 1108,
      "manually creating": 19566,
      "technique achieves": 32615,
      "edit distance": 8823,
      "false positive": 10959,
      "positive rate": 25197,
      "prediction problems": 25433,
      "exponentially large": 10630,
      "large context": 16936,
      "generate set": 12324,
      "sentence used": 29547,
      "time instead": 33128,
      "dl applications": 8494,
      "edge cases": 8819,
      "challenging domains": 4383,
      "observed language": 23239,
      "modeling long": 20898,
      "technique applied": 32616,
      "address questions": 1355,
      "spanning 1000": 30597,
      "transformer framework": 33717,
      "detection systems": 7882,
      "approach implementing": 2295,
      "components including": 5313,
      "common transformer": 5018,
      "inference training": 14819,
      "effect context": 8852,
      "neural scaling": 22758,
      "models single": 21982,
      "provide final": 26702,
      "daily life": 6578,
      "suggestions real": 31588,
      "pass1 humaneval": 24374,
      "9b tokens": 612,
      "llms brought": 18453,
      "models ensuring": 21209,
      "power edge": 25319,
      "techniques using": 32671,
      "rate compared": 27264,
      "raise concerns": 27163,
      "interpretability llms": 15458,
      "llms simply": 18953,
      "methods effectiveness": 20024,
      "jailbreak prompts": 15697,
      "llms seen": 18929,
      "quantify severity": 26987,
      "different demographics": 8070,
      "analysis provide": 1951,
      "insights choice": 15069,
      "social groups": 30422,
      "unprecedented performance": 34420,
      "especially visual": 9750,
      "inputs enabling": 15045,
      "successfully evade": 31538,
      "dl models": 8495,
      "data applying": 6600,
      "support limited": 31710,
      "substantial number": 31470,
      "lack clear": 15977,
      "methods identifying": 20045,
      "tremendous advances": 33882,
      "introduce study": 15534,
      "adversarial training": 1519,
      "models unlike": 22095,
      "pretrained extensive": 25642,
      "finetuning paper": 11468,
      "domain code": 8556,
      "closed llms": 4677,
      "vary lot": 35166,
      "context significantly": 5918,
      "user requirements": 34669,
      "poses security": 25172,
      "identify model": 14014,
      "utilizes techniques": 34967,
      "enhancing decisionmaking": 9559,
      "especially complex": 9731,
      "accurate identification": 923,
      "llms raised": 18884,
      "vast quantities": 35190,
      "designed empower": 7727,
      "currently supports": 6545,
      "better efficiency": 3600,
      "datasets downstream": 7099,
      "demonstrating strong": 7589,
      "strong generalizability": 31172,
      "higher precision": 13601,
      "large transformers": 17283,
      "inference transformer": 14820,
      "enable fast": 9287,
      "inference framework": 14779,
      "framework designs": 11839,
      "similar accuracy": 30096,
      "information detection": 14859,
      "including bart": 14459,
      "probability model": 25972,
      "instead relying": 15125,
      "relying manual": 27978,
      "bard claude": 3096,
      "model detect": 20465,
      "models opt13b": 21777,
      "statistical power": 31020,
      "llama7b models": 18235,
      "reliably detect": 27957,
      "unseen apis": 34433,
      "generation study": 12606,
      "diverse complex": 8416,
      "performance generating": 24616,
      "generating entire": 12421,
      "strategy best": 31115,
      "understanding long": 34249,
      "long instructions": 19174,
      "data attribution": 6604,
      "european union": 9811,
      "90 performance": 590,
      "performance scales": 24745,
      "size results": 30281,
      "build high": 3912,
      "leveraging natural": 17790,
      "changes human": 4410,
      "generation instead": 12523,
      "predictions training": 25445,
      "attention ability": 2712,
      "accurate inference": 924,
      "llms serves": 18932,
      "llms accuracy": 18408,
      "dramatically improved": 8718,
      "structure inference": 31213,
      "furthermore use": 12014,
      "pretraining scratch": 25836,
      "leverages capabilities": 17766,
      "make possible": 19477,
      "control group": 6051,
      "detection language": 7868,
      "finally make": 11198,
      "contexts zeroshot": 5947,
      "70b code": 504,
      "7b outperforms": 538,
      "opensource software": 23544,
      "weights large": 35508,
      "data domain": 6685,
      "llama evaluate": 18095,
      "additionally discuss": 1281,
      "offensive content": 23283,
      "false positives": 10961,
      "correctly detected": 6203,
      "increasing need": 14625,
      "gpt4 furthermore": 13078,
      "effectively enhance": 8916,
      "explore types": 10602,
      "work large": 35731,
      "novel geometric": 23084,
      "geometric perspective": 12727,
      "adversarial prompting": 1517,
      "safety guardrails": 29047,
      "maintaining good": 19424,
      "performance safe": 24744,
      "gradient information": 13188,
      "temporal resolution": 32700,
      "llms hpc": 18677,
      "gpt3 generate": 12994,
      "original gpt3": 23706,
      "competitive superior": 5231,
      "llama27b models": 18210,
      "paper raise": 24127,
      "selection strategy": 29396,
      "templates high": 32692,
      "success rates": 31526,
      "qualitative approach": 26932,
      "llm significant": 18367,
      "engineering instruction": 9466,
      "performance assessment": 24526,
      "popular parameterefficient": 25134,
      "consistently activate": 5745,
      "attacks necessary": 2694,
      "threat model": 33090,
      "editing methods": 8828,
      "editing method": 8827,
      "effective defense": 8866,
      "incident response": 14445,
      "approach addressing": 2233,
      "unclear paper": 34128,
      "similar written": 30121,
      "increase maximum": 14597,
      "efficiently produce": 9073,
      "evaluate technique": 9867,
      "data identify": 6728,
      "costs llm": 6271,
      "novel algorithms": 23054,
      "expanding context": 10340,
      "approaches identify": 2375,
      "auc score": 2772,
      "harmful language": 13442,
      "sacrificing model": 29026,
      "content address": 5849,
      "provable guarantees": 26666,
      "queries existing": 27020,
      "available following": 2976,
      "optimizing large": 23656,
      "harmful behaviors": 13440,
      "research reveals": 28359,
      "advocate research": 1531,
      "research efforts": 28310,
      "applications involving": 2158,
      "llms generated": 18636,
      "smaller opensourced": 30393,
      "chatgpt performance": 4493,
      "showcase capability": 29835,
      "stateoftheart proprietary": 30979,
      "text inputs": 32899,
      "extremely simple": 10802,
      "strategies including": 31108,
      "cost finally": 6244,
      "effective alignment": 8859,
      "effective natural": 8889,
      "accuracy degradation": 867,
      "groups work": 13306,
      "alignment models": 1774,
      "given candidate": 12741,
      "llama glm": 18107,
      "challenge training": 4331,
      "preserves data": 25605,
      "malicious usage": 19524,
      "model quantized": 20742,
      "method provide": 19961,
      "model applications": 20370,
      "better code": 3597,
      "testing allows": 32800,
      "stochastic parrots": 31071,
      "llms available": 18442,
      "lms based": 19070,
      "based approximate": 3137,
      "data popular": 6798,
      "remedy issue": 28070,
      "task description": 32108,
      "changing semantic": 4416,
      "predictions language": 25441,
      "way evaluate": 35430,
      "model way": 20863,
      "biases model": 3679,
      "particularly resourceconstrained": 24355,
      "generative process": 12703,
      "security privacy": 29349,
      "layer outputs": 17428,
      "finetuning specifically": 11533,
      "strategy use": 31129,
      "finetuning result": 11511,
      "extra parameters": 10742,
      "methods paramount": 20073,
      "dataset real": 7029,
      "models grant": 21336,
      "model weight": 20864,
      "step development": 31040,
      "collection instruction": 4930,
      "finetuning public": 11500,
      "lora efficient": 19229,
      "general performance": 12182,
      "models considerable": 21096,
      "models meta": 21718,
      "high computation": 13556,
      "method termed": 19979,
      "investigate persona": 15591,
      "solution code": 30471,
      "better ranking": 3622,
      "new stateofthearts": 22852,
      "automatically detect": 2909,
      "proposed technique": 26621,
      "achieving 70": 1081,
      "improves llm": 14381,
      "llm size": 18369,
      "size increases": 30252,
      "aims democratize": 1661,
      "access llms": 821,
      "associated large": 2646,
      "additional modality": 1260,
      "instead feeding": 15121,
      "manually review": 19572,
      "generated textual": 12395,
      "generated chatgpt": 12346,
      "capabilities field": 4018,
      "attacks introduce": 2691,
      "distribution consequently": 8390,
      "task conduct": 32098,
      "effectiveness transferability": 8969,
      "examples generated": 10125,
      "contributing success": 6035,
      "deploying models": 7643,
      "outcomes underscore": 23736,
      "models superior": 22030,
      "models undergone": 22092,
      "undergone instruction": 34148,
      "need knowledge": 22633,
      "reveal various": 28807,
      "detection classification": 7863,
      "detection rate": 7878,
      "remarkably low": 28066,
      "detection study": 7881,
      "employing large": 9258,
      "maintaining models": 19427,
      "models log": 21686,
      "demonstrate used": 7510,
      "evaluation analysis": 9918,
      "novel study": 23112,
      "use mechanistic": 34548,
      "mechanistic interpretability": 19756,
      "interpretability approaches": 15457,
      "heads layers": 13481,
      "model prior": 20726,
      "easily detected": 8795,
      "toxicity classifiers": 33318,
      "models grown": 21344,
      "text humanauthored": 32892,
      "framework work": 11906,
      "tasks practical": 32452,
      "internal representations": 15441,
      "provide practical": 26722,
      "multihead attention": 22285,
      "geometric interpretation": 12726,
      "theoretical results": 33051,
      "emerged dominant": 9156,
      "problematic model": 26021,
      "alternative methods": 1855,
      "growing demand": 13314,
      "high levels": 13573,
      "gpt large": 12850,
      "study tendency": 31402,
      "evaluation pipeline": 9988,
      "lightweight language": 17840,
      "resourceconstrained scenarios": 28421,
      "findings design": 11234,
      "showcases potential": 29846,
      "volume demonstrates": 35387,
      "allows customization": 1811,
      "input making": 15017,
      "summarization incontext": 31613,
      "neuron level": 22767,
      "causal effect": 4237,
      "analysis tools": 1972,
      "tools require": 33274,
      "mitigate hallucinations": 20253,
      "designed adversarial": 7720,
      "able successfully": 755,
      "surpasses traditional": 31751,
      "presents comprehensive": 25577,
      "evaluates models": 9889,
      "parameters like": 24267,
      "popularity ability": 25141,
      "scraped internet": 29288,
      "extent phenomenon": 10724,
      "order build": 23670,
      "surge popularity": 31729,
      "benchmark future": 3389,
      "users professional": 34699,
      "greatly reduce": 13275,
      "framework incorporates": 11868,
      "incorporates innovative": 14572,
      "innovative techniques": 15000,
      "process employed": 26058,
      "assigning higher": 2626,
      "robust measurement": 28935,
      "greater resilience": 13270,
      "users preferences": 34698,
      "making difficult": 19501,
      "encoderdecoder transformer": 9373,
      "various coderelated": 35080,
      "finetuning schemes": 11518,
      "llama 34b": 18064,
      "framework mitigating": 11883,
      "method finetuning": 19922,
      "insight design": 15066,
      "algorithm use": 1715,
      "step size": 31048,
      "evaluations llms": 10033,
      "systems large": 31904,
      "enhance generation": 9513,
      "potential llm": 25273,
      "complete test": 5251,
      "code llama34b": 4774,
      "llama34b model": 18227,
      "improved generation": 14311,
      "automatically effectively": 2910,
      "methods especially": 20026,
      "based incontext": 3175,
      "encompasses types": 9388,
      "preserving models": 25610,
      "science research": 29245,
      "significantly increases": 30065,
      "generation open": 12564,
      "vicuna chatglm": 35248,
      "maintain general": 19415,
      "issues associated": 15666,
      "requirements existing": 28237,
      "small large": 30351,
      "resource efficiency": 28411,
      "using scoring": 34907,
      "study establishes": 31325,
      "traditional evaluation": 33345,
      "wider array": 35584,
      "massive size": 19629,
      "minimal computation": 20183,
      "inference maintaining": 14791,
      "inference context": 14769,
      "computational savings": 5481,
      "layers model": 17442,
      "aims produce": 1670,
      "reduced computation": 27729,
      "level model": 17734,
      "chat vicuna": 4446,
      "input changes": 15004,
      "performance extensive": 24594,
      "strategies llms": 31110,
      "capabilities exhibit": 4014,
      "severe threat": 29753,
      "llms triggered": 19011,
      "investigate recent": 15596,
      "line work": 17983,
      "llms probability": 18869,
      "longer ones": 19199,
      "role predicting": 28960,
      "short sequences": 29815,
      "restricted extensive": 28532,
      "surpasses existing": 31744,
      "knowledge sharing": 15905,
      "effectively replace": 8931,
      "texts specific": 33001,
      "potential downstream": 25254,
      "detection recent": 7879,
      "focus generative": 11648,
      "innovative strategy": 14999,
      "strategy designed": 31117,
      "effectiveness extensive": 8944,
      "hardware constraints": 13428,
      "memory paper": 19825,
      "content particularly": 5869,
      "gap conducted": 12087,
      "implemented finetuning": 14168,
      "detection paper": 7874,
      "accelerate training": 794,
      "training regimes": 33598,
      "finetuning stateoftheart": 11536,
      "optimizing training": 23657,
      "ai alignment": 1596,
      "models reducing": 21915,
      "transparent ai": 33864,
      "secure efficient": 29345,
      "context extrapolation": 5890,
      "despite advantages": 7770,
      "version original": 35232,
      "phase results": 24918,
      "responses detecting": 28489,
      "core functionalities": 6152,
      "domain poses": 8582,
      "developed specialized": 7934,
      "approach exhibits": 2274,
      "instructblip mplugowl2": 15133,
      "reveal connection": 28793,
      "automated generation": 2862,
      "higher scores": 13604,
      "additionally analyze": 1272,
      "requirements limited": 28240,
      "problem use": 26017,
      "growing concerns": 13313,
      "particular use": 24344,
      "datasets provides": 7164,
      "cover 30": 6315,
      "relative ease": 27878,
      "presence absence": 25508,
      "domain questions": 8586,
      "achieve carefully": 950,
      "human machine": 13844,
      "machine intelligence": 19342,
      "chatgpt field": 4468,
      "mistral mixtral": 20233,
      "level work": 17737,
      "sufficient achieve": 31560,
      "commonly executed": 5025,
      "images sharing": 14087,
      "diverse new": 8445,
      "authorship identification": 2839,
      "online authorship": 23360,
      "orders magnitudes": 23684,
      "popularity recent": 25146,
      "strong simple": 31191,
      "techniques proposed": 32657,
      "enable comprehensive": 9285,
      "fairness llms": 10927,
      "predicting judgments": 25418,
      "societal biases": 30437,
      "predictions study": 25444,
      "legal tasks": 17698,
      "social factors": 30421,
      "indicate proposed": 14694,
      "legal sector": 17697,
      "specialised legal": 30659,
      "potential method": 25276,
      "method mitigate": 19946,
      "mitigate bias": 20248,
      "sampled data": 29068,
      "correlation training": 6220,
      "data adversarial": 6595,
      "additional results": 1264,
      "empirical analyses": 9218,
      "rate features": 27265,
      "art form": 2513,
      "gpt4 gemini": 13079,
      "observation develop": 23218,
      "effective prompting": 8893,
      "success language": 31513,
      "roberta llama2": 28920,
      "federated finetuning": 11049,
      "service platform": 29658,
      "attention potential": 2735,
      "rich insights": 28874,
      "shown models": 29901,
      "ensure safe": 9607,
      "llama1 llama2": 18154,
      "efficiency traditional": 9017,
      "creating large": 6368,
      "llms attention": 18437,
      "using fewshot": 34777,
      "patterns including": 24411,
      "research llm": 28334,
      "important milestone": 14205,
      "use specific": 34571,
      "demonstrate average": 7437,
      "focal point": 11643,
      "potential significantly": 25297,
      "2023 using": 236,
      "using latest": 34817,
      "inputs code": 15044,
      "examples making": 10135,
      "method practical": 19956,
      "harming performance": 13446,
      "sophisticated methods": 30525,
      "techniques targeted": 32666,
      "compared gradientbased": 5139,
      "nvidia rtx": 23196,
      "48gb gpu": 408,
      "accelerate research": 793,
      "codebase publicly": 4832,
      "work suggests": 35793,
      "models conducted": 21092,
      "techniques significantly": 32665,
      "chatgpt gemini": 4471,
      "help understand": 13514,
      "novel tool": 23120,
      "llama study": 18146,
      "suit specific": 31594,
      "data verbatim": 6908,
      "size scales": 30282,
      "quantized large": 27013,
      "maintaining model": 19426,
      "llama2 families": 18175,
      "exhibit exceptional": 10217,
      "7b instruct": 534,
      "4x larger": 417,
      "3b 7b": 365,
      "llm benchmarks": 18277,
      "mistral7b datasets": 20238,
      "datasets sst2": 7174,
      "advanced training": 1443,
      "loss landscape": 19246,
      "detection strategy": 7880,
      "constructing prompts": 5809,
      "reduce time": 27726,
      "time cost": 33117,
      "information related": 14903,
      "work perform": 35746,
      "aims investigate": 1669,
      "claude llama": 4636,
      "floatingpoint operations": 11624,
      "semantic similarities": 29473,
      "using fl": 34782,
      "resulting lower": 28556,
      "research highlighted": 28321,
      "inference prompt": 14802,
      "understanding semantics": 34269,
      "applications document": 2150,
      "llms uncover": 19015,
      "important tool": 14214,
      "concerns potential": 5543,
      "potential misuse": 25277,
      "distribution gap": 8392,
      "languages findings": 16876,
      "effective learning": 8881,
      "achieves relative": 1059,
      "relative improvements": 27881,
      "validated using": 35000,
      "llm practitioners": 18349,
      "develop effective": 7914,
      "attacks llm": 2692,
      "novel trainingfree": 23123,
      "services train": 29662,
      "types training": 34073,
      "nvidia gpus": 23194,
      "achieve exact": 962,
      "exact training": 10091,
      "significantly decreases": 30043,
      "systematic empirical": 31869,
      "pro vision": 25963,
      "secondorder information": 29336,
      "llama gemini": 18105,
      "information hessian": 14871,
      "datasets case": 7072,
      "length target": 17712,
      "complex scientific": 5293,
      "data need": 6783,
      "generation uses": 12630,
      "surpasses current": 31742,
      "efficiency quality": 9012,
      "schemes mitigate": 29235,
      "study llama": 31357,
      "furthermore previous": 12009,
      "illustrate llms": 14046,
      "technique empowers": 32620,
      "model autonomously": 20382,
      "considerable size": 5710,
      "strategies experiments": 31105,
      "prompts evaluate": 26415,
      "llms revealing": 18921,
      "limitations proposed": 17933,
      "billions data": 3727,
      "work suggest": 35792,
      "various sources": 35145,
      "accuracy specific": 912,
      "identify data": 14008,
      "significantly boosting": 30039,
      "groupedquery attention": 13300,
      "rotary positional": 28974,
      "multiple language": 22396,
      "foundational capabilities": 11802,
      "plms downstream": 25042,
      "using fixed": 34781,
      "work consider": 35680,
      "llms detecting": 18535,
      "api usage": 2103,
      "class data": 4582,
      "approach termed": 2346,
      "analysis revealing": 1958,
      "largescale deep learning": 17349,
      "recent years witnessed": 27584,
      "bert gpt2 xlnet": 3513,
      "parameter language models": 24186,
      "automatic code generation": 2877,
      "different language models": 8089,
      "representations bert gpt2": 28158,
      "neural network dnn": 22742,
      "network dnn models": 22689,
      "modern machine learning": 22167,
      "unlike prior work": 34401,
      "gpt natural language": 12859,
      "transformer gpt2 model": 33721,
      "using pretrained t5": 34882,
      "transformer based language": 33707,
      "code generation model": 4761,
      "proposed method effectively": 26606,
      "samples training set": 29089,
      "samples language models": 29079,
      "shown large pretrained": 29897,
      "models llms bert": 21485,
      "model compression propose": 20435,
      "matches exceeds performance": 19651,
      "language processing applications": 16775,
      "applications use large": 2181,
      "pretrained models work": 25730,
      "semantic meaning original": 29461,
      "language models nlms": 16633,
      "identifiable information pii": 13995,
      "language model api": 16117,
      "open pretrained transformer": 23412,
      "language models advance": 16243,
      "task existing methods": 32118,
      "language model data": 16130,
      "false positive rate": 10960,
      "issues paper propose": 15670,
      "training time instead": 33634,
      "language modeling long": 16219,
      "modeling long text": 20899,
      "address questions introduce": 1356,
      "novel approach implementing": 23061,
      "neural scaling laws": 22759,
      "llms propose novel": 18876,
      "models llms seen": 21626,
      "conduct empirical study": 5599,
      "data public httpsgithubcomnlpxucanwizardlm": 6819,
      "recently gained attention": 27599,
      "pretrained t5 model": 25752,
      "competitive performance compared": 5224,
      "models llms raised": 21610,
      "models trained vast": 22076,
      "various tasks require": 35151,
      "inference transformer models": 14821,
      "transformer models using": 33734,
      "generation address issue": 12453,
      "demonstrate proposed method": 7484,
      "language models opt13b": 16640,
      "prompt chatgpt generate": 26312,
      "presents unique challenges": 25600,
      "build high quality": 3913,
      "models finetuning large": 21280,
      "highquality text generation": 13702,
      "language models mbert": 16617,
      "predictions training data": 25446,
      "safety alignment llms": 29043,
      "covering wide range": 6329,
      "lowresource languages using": 19315,
      "models parameterefficient finetuning": 21802,
      "framework leverages capabilities": 11879,
      "stateoftheart performance open": 30973,
      "performance open models": 24701,
      "7b outperforms llama": 539,
      "work large language": 35732,
      "novel geometric perspective": 23085,
      "instructions training large": 15278,
      "paper raise concerns": 24128,
      "models using small": 22105,
      "engineering instruction tuning": 9467,
      "popular parameterefficient finetuning": 25135,
      "model editing methods": 20479,
      "similar written humans": 30122,
      "publicly available following": 26856,
      "llms inference time": 18719,
      "findings suggest finetuning": 11257,
      "advocate research efforts": 1532,
      "high inference costs": 13570,
      "stateoftheart proprietary models": 30980,
      "performance compared previous": 24552,
      "effective natural language": 8890,
      "align human values": 1727,
      "learning code generation": 17553,
      "parameterefficient finetuning methods": 24208,
      "language models deployed": 16303,
      "language model weights": 16211,
      "language model applications": 16118,
      "code completion tasks": 4727,
      "changing semantic meaning": 4417,
      "success rate compared": 31525,
      "language models grant": 16375,
      "access model weights": 823,
      "lora efficient finetuning": 19230,
      "models sizes 7b": 21985,
      "language models meta": 16620,
      "downstream tasks compared": 8689,
      "language models specialized": 16709,
      "despite remarkable success": 7810,
      "remarkable success various": 28062,
      "high computation cost": 13557,
      "solution code generation": 30472,
      "associated large language": 2647,
      "hope work contribute": 13758,
      "different aspects including": 8051,
      "undergone instruction tuning": 34149,
      "examine impact various": 10102,
      "employing large language": 9259,
      "llms different architectures": 18540,
      "presents novel study": 25590,
      "use mechanistic interpretability": 34549,
      "model prior knowledge": 20727,
      "various tasks particularly": 35150,
      "present novel solution": 25548,
      "gpt large language": 12851,
      "llms automatically generate": 18441,
      "work propose framework": 35753,
      "nvidia a100 gpu": 23193,
      "paper presents comprehensive": 24096,
      "models llms generation": 21540,
      "ability generate humanlike": 675,
      "training costs paper": 33462,
      "incorporates innovative techniques": 14573,
      "llama mistral models": 18127,
      "opensource models code": 23530,
      "systems large language": 31905,
      "code llama34b model": 4775,
      "llms llama chatgpt": 18768,
      "based incontext learning": 3176,
      "experimental results language": 10400,
      "models ranging size": 21891,
      "parameters demonstrate effectiveness": 24238,
      "social science research": 30432,
      "massive size poses": 19630,
      "hindering widespread adoption": 13719,
      "maintaining models performance": 19428,
      "llama2 chat vicuna": 18167,
      "model performance extensive": 20698,
      "performance extensive experiments": 24595,
      "analysis using large": 1977,
      "models llms triggered": 21653,
      "paper investigate recent": 24071,
      "language models specific": 16710,
      "detection paper presents": 7875,
      "despite widespread use": 7829,
      "llava instructblip mplugowl2": 18246,
      "current stateoftheart methods": 6534,
      "outperforming existing approaches": 23797,
      "experiments various llms": 10500,
      "strong simple baseline": 31192,
      "training data paper": 33486,
      "strong correlation training": 31168,
      "transformer models bert": 33732,
      "gpt35 gpt4 gemini": 13022,
      "prior work shown": 25942,
      "work shown models": 35785,
      "single nvidia rtx": 30218,
      "codebase publicly available": 4833,
      "quantized large language": 27014,
      "models opt llama2": 21776,
      "llms exhibit exceptional": 18580,
      "mistral 7b instruct": 20229,
      "select highquality data": 29379,
      "advanced training techniques": 1444,
      "bard claude llama": 3097,
      "concerns potential misuse": 5544,
      "methods primarily focus": 20078,
      "language models generated": 16360,
      "models paper study": 21798,
      "conduct systematic empirical": 5624,
      "gemini pro vision": 12143,
      "like chatgpt llama": 17855,
      "enhance ability large": 9502,
      "surpasses current stateoftheart": 31743,
      "generation using llms": 12634,
      "rotary positional embedding": 28975,
      "plms downstream tasks": 25043,
      "novel approach termed": 23063,
      "deep neural network dnn": 7339,
      "neural network dnn models": 22743,
      "pretrained transformer gpt2 model": 25760,
      "transformer based language models": 33708,
      "language models including gpt2": 16387,
      "large pretrained transformer models": 17269,
      "shown large pretrained language": 29898,
      "language models llms bert": 16434,
      "natural language processing applications": 22536,
      "language modeling long text": 16220,
      "language models llms seen": 16558,
      "language models llms raised": 16545,
      "results demonstrate proposed method": 28592,
      "demonstrate proposed method yields": 7485,
      "various text generation models": 35155,
      "large language models capable": 16997,
      "large language models finetuning": 17023,
      "language models finetuning large": 16346,
      "models finetuning large language": 21281,
      "stateoftheart performance open models": 30974,
      "work large language models": 35733,
      "instructions training large language": 15279,
      "code publicly available following": 4804,
      "using parameterefficient finetuning methods": 34871,
      "models sizes 7b 13b": 21986,
      "large language models meta": 17172,
      "associated large language models": 2648,
      "paper presents novel study": 24105,
      "propose reinforcement learning rl": 26563,
      "gpt large language model": 12852,
      "conduct extensive experiments various": 5609,
      "paper introduces novel approach": 24065,
      "language models llms generation": 16481,
      "ability generate humanlike text": 676,
      "high training costs paper": 13588,
      "systems large language models": 31906,
      "analysis using large language": 1978,
      "language models llms triggered": 16582,
      "extensive experiments various llms": 10706,
      "quantized large language models": 27015,
      "text large language models": 32905,
      "large language models generated": 17033,
      "deep neural network dnn models": 7340,
      "generative pretrained transformer gpt2 model": 12696,
      "shown large pretrained language models": 29899,
      "generation large language models demonstrated": 12536,
      "modern large language models llms": 22164,
      "popularity large language models llms": 25145,
      "large language models llms seen": 17143,
      "large language models llms raised": 17135,
      "large language models finetuning large": 17024,
      "language models finetuning large language": 16347,
      "models finetuning large language models": 21282,
      "instructions training large language models": 15280,
      "framework large language models large": 11877,
      "associated large language models llms": 2649,
      "large language models llms generation": 17090,
      "large language models llms triggered": 17158,
      "text large language models llms": 32906,
      "visualizing": 35364,
      "recurring": 27686,
      "converging": 6088,
      "stones": 31072,
      "approximates": 2416,
      "16x": 168,
      "tokenfree": 33208,
      "synchronous": 31806,
      "diverges": 8408,
      "businesses": 3945,
      "exacerbated": 10085,
      "formulates": 11774,
      "trading": 33340,
      "reads": 27307,
      "receptive": 27628,
      "gpt1": 12865,
      "attributing": 2767,
      "song": 30521,
      "asymmetry": 2676,
      "stitch": 31066,
      "reciprocal": 27633,
      "dropout": 8743,
      "resembles": 28384,
      "ssl": 30844,
      "systemic": 31884,
      "precisions": 25395,
      "sensitivitybased": 29523,
      "plateau": 25007,
      "703": 501,
      "saved": 29114,
      "goto": 12835,
      "absorbed": 766,
      "fulllength": 11946,
      "rubert": 29003,
      "rugpt3": 29005,
      "unrealistic": 34424,
      "trainingbased": 33646,
      "262": 280,
      "weighed": 35492,
      "anticipated": 2093,
      "equivariant": 9690,
      "intersectionality": 15472,
      "en": 9278,
      "personalised": 24882,
      "closesourced": 4698,
      "increment": 14654,
      "jobs": 15708,
      "locality": 19137,
      "colab": 4898,
      "flash": 11608,
      "tutorial": 34032,
      "born": 3833,
      "underestimate": 34139,
      "timeseries": 33168,
      "forecast": 11721,
      "mimicry": 20172,
      "attributions": 2769,
      "compatibility": 5202,
      "transcends": 33663,
      "522": 435,
      "untrained": 34462,
      "interchunk": 15399,
      "compounded": 5332,
      "present opensource": 25549,
      "model bias": 20401,
      "attention transformer": 2742,
      "attention multiple": 2728,
      "gpt2 present": 12936,
      "larger dataset": 17320,
      "tens hundreds": 32718,
      "test loss": 32773,
      "pipeline model": 24969,
      "datasets bert": 7071,
      "requiring model": 28275,
      "machine authors": 19340,
      "model fewer": 20518,
      "tasks sequence": 32495,
      "work addresses": 35666,
      "networks like": 22713,
      "different attention": 8052,
      "size neural": 30266,
      "models fit": 21285,
      "experiments compared": 10425,
      "does contain": 8525,
      "sacrificing accuracy": 29025,
      "trained specific": 33428,
      "using transfer": 34931,
      "corresponding word": 6231,
      "models important": 21370,
      "roberta deberta": 28916,
      "goal paper": 12807,
      "model single": 20787,
      "evaluate endtoend": 9835,
      "expensive work": 10370,
      "memory model": 19822,
      "evaluating model": 9906,
      "uses decoder": 34708,
      "proxy perplexity": 26800,
      "highly correlates": 13661,
      "attention scores": 2741,
      "scores subsequent": 29282,
      "network architecture": 22684,
      "tasks word": 32552,
      "method reduce": 19964,
      "models simultaneously": 21981,
      "provides significant": 26762,
      "transformers generate": 33781,
      "code runs": 4815,
      "memory capacity": 19804,
      "metrics paper": 20145,
      "language identification": 16093,
      "task numerous": 32168,
      "llms excellent": 18577,
      "access large": 819,
      "design learning": 7708,
      "decoding enhance": 7272,
      "depends number": 7625,
      "llms determine": 18536,
      "early training": 8782,
      "model short": 20778,
      "evergrowing size": 10051,
      "flant5 outperform": 11601,
      "models hierarchical": 21352,
      "constraints aggregating": 5789,
      "approaches data": 2371,
      "require training": 28227,
      "models makes": 21698,
      "presents challenges": 25575,
      "paper identify": 24057,
      "studies including": 31272,
      "reducing gender": 27748,
      "human life": 13841,
      "superiority approach": 31662,
      "sequences training": 29617,
      "semantic role": 29469,
      "role labeling": 28958,
      "tasks addition": 32234,
      "limitations using": 17935,
      "evaluating instruction": 9896,
      "crucial problem": 6444,
      "acquired knowledge": 1121,
      "llms necessitates": 18814,
      "scenarios tested": 29222,
      "models practice": 21829,
      "adaptability llms": 1171,
      "various generation": 35100,
      "framework efficiently": 11842,
      "qlora efficient": 26923,
      "reduce average": 27700,
      "analysis instruction": 1930,
      "appropriate responses": 2408,
      "data identifying": 6729,
      "identifying relevant": 14023,
      "diverse instruction": 8433,
      "method successfully": 19977,
      "preserves original": 25606,
      "fluent samples": 11637,
      "unconditional zeroshot": 34132,
      "crucial comprehend": 6439,
      "learning ssl": 17656,
      "evaluations interestingly": 10032,
      "close gap": 4672,
      "terms f1score": 32746,
      "come cost": 4966,
      "cost significant": 6254,
      "writing large": 35852,
      "strongly correlates": 31208,
      "transformer recent": 33739,
      "model linear": 20617,
      "particular introduce": 24339,
      "opt125m model": 23603,
      "suggest large": 31573,
      "tokenbytoken generation": 33207,
      "readily applied": 27301,
      "metric used": 20128,
      "rigorous research": 28885,
      "irrespective models": 15646,
      "like knowledge": 17878,
      "based gptj": 3171,
      "using activation": 34726,
      "efficient solution": 9059,
      "retraining scratch": 28728,
      "effect different": 8853,
      "pretrained pile": 25742,
      "experiments pythia": 10474,
      "rates various": 27271,
      "humanwritten text": 13938,
      "bias gender": 3645,
      "times higher": 33162,
      "effective control": 8862,
      "considering different": 5723,
      "improvements code": 14356,
      "make difficult": 19467,
      "largescale datasets": 17347,
      "methods context": 20013,
      "design particular": 7713,
      "test methods": 32775,
      "using longer": 34827,
      "tool identify": 33259,
      "gpts llama": 13164,
      "challenges opportunities": 4364,
      "necessity finetuning": 22615,
      "pretrained llama": 25703,
      "large bias": 16932,
      "individual layers": 14716,
      "techniques improving": 32643,
      "achieving acceptable": 1082,
      "accessible models": 833,
      "study identifies": 31341,
      "lower temperature": 19291,
      "range 05": 27186,
      "contextual embeddings": 5949,
      "understanding transformers": 34277,
      "does harm": 8528,
      "bias terms": 3664,
      "introduce experimental": 15506,
      "impact important": 14125,
      "approach extends": 2278,
      "previous tokens": 25890,
      "attention recent": 2738,
      "million tokens": 20166,
      "tokens generated": 33229,
      "opt pythia": 23601,
      "deployment largescale": 7654,
      "modeling pairwise": 20903,
      "research aimed": 28288,
      "interpretable attention": 15461,
      "weights llm": 35510,
      "efficient adaptation": 9021,
      "learning learn": 17597,
      "order understand": 23678,
      "questions demonstrating": 27105,
      "llms retrieval": 18919,
      "retrieval significantly": 28753,
      "model retrievalaugmented": 20760,
      "good bad": 12817,
      "transformers increasing": 33785,
      "zeroshot language": 35982,
      "llms uncovering": 19016,
      "benchmarks demonstrates": 3438,
      "exhibits remarkable": 10249,
      "work observe": 35737,
      "observe finetuning": 23227,
      "accuracy zeroshot": 916,
      "models highlight": 21355,
      "works make": 35816,
      "behavior icl": 3314,
      "llms costly": 18495,
      "takes account": 32031,
      "scheme designed": 29233,
      "different groups": 8080,
      "time sparsity": 33143,
      "improves efficiency": 14375,
      "holistically evaluate": 13745,
      "contributes improving": 6033,
      "step en": 31042,
      "en route": 9279,
      "route enabling": 28989,
      "enabling widespread": 9330,
      "lms crucial": 19076,
      "detecting mitigating": 7858,
      "trainingbased methods": 33647,
      "output logits": 23874,
      "reducing average": 27743,
      "samples limited": 29083,
      "applicability approach": 2118,
      "embeddings capture": 9138,
      "scaling findings": 29162,
      "balance model": 3079,
      "distributed model": 8386,
      "7b chat": 533,
      "trained direct": 33390,
      "average including": 3020,
      "small diverse": 30340,
      "domains recent": 8635,
      "methods test": 20103,
      "reducing llm": 27753,
      "new challenges": 22786,
      "adapting new": 1213,
      "llm large": 18327,
      "reason introduce": 27356,
      "accuracy evaluating": 875,
      "effectively paper": 8926,
      "llms implement": 18683,
      "comparable existing": 5078,
      "enabling model": 9325,
      "frontier large": 11932,
      "extend model": 10656,
      "neurons consistently": 22769,
      "understanding mechanisms": 34251,
      "models contextual": 21105,
      "performance spatial": 24761,
      "study automatic": 31303,
      "work delve": 35689,
      "models selected": 21959,
      "explore behavior": 10574,
      "google colab": 12828,
      "increasingly large": 14640,
      "method prune": 19962,
      "big challenge": 3699,
      "size context": 30242,
      "models fields": 21269,
      "domain natural": 8575,
      "domains llms": 8631,
      "finetuning terms": 11547,
      "identify strong": 14017,
      "techniques terms": 32667,
      "approach creating": 2254,
      "presents scalable": 25593,
      "prominent language": 26266,
      "models diverge": 21163,
      "preserve model": 25603,
      "sample complexity": 29065,
      "set trained": 29712,
      "incorporating novel": 14578,
      "novel methods": 23099,
      "providing additional": 26771,
      "projection weight": 26252,
      "potential improving": 25265,
      "exciting promise": 10184,
      "analysis models": 1943,
      "techniques approaches": 32628,
      "limitations stateoftheart": 17934,
      "understanding latent": 34243,
      "provide opensource": 26720,
      "sizes existing": 30298,
      "models means": 21710,
      "adaptation technique": 1194,
      "families roberta": 10971,
      "hallucinations paper": 13398,
      "model limited": 20616,
      "financial applications": 11217,
      "llms computationally": 18483,
      "lora qlora": 19234,
      "llms financial": 18607,
      "tasks financial": 32331,
      "financial tasks": 11222,
      "surge large": 31725,
      "original distribution": 23704,
      "llms fewer": 18604,
      "uses features": 34709,
      "instruction pairs": 15173,
      "pretraining ultimately": 25850,
      "achieve specific": 989,
      "arbitrary batch": 2429,
      "finetuning required": 11509,
      "specific subset": 30716,
      "method address": 19872,
      "complex structure": 5294,
      "approaches lead": 2379,
      "framework experimental": 11852,
      "models combinatorial": 21062,
      "stateoftheart results natural": 30984,
      "pretrained transformer language": 25762,
      "different attention heads": 8053,
      "models trained specific": 22075,
      "using transfer learning": 34932,
      "code data used": 4741,
      "popular pretrained language": 25137,
      "models gpt2 model": 21326,
      "long sequence lengths": 19177,
      "evaluate endtoend performance": 9836,
      "remains unclear paper": 28017,
      "neural network architecture": 22739,
      "language models downstream": 16312,
      "model pretrained language": 20721,
      "depends number parameters": 7626,
      "zeroshot fewshot chainofthought": 35972,
      "reducing gender bias": 27749,
      "demonstrate superiority approach": 7506,
      "instruction finetuned language": 15155,
      "semantic role labeling": 29470,
      "models llms necessitates": 21590,
      "tasks findings reveal": 32333,
      "selfsupervised learning ssl": 29438,
      "existing methods require": 10294,
      "come cost significant": 4967,
      "modern transformer models": 22174,
      "llms llama vicuna": 18773,
      "better alignment human": 3590,
      "techniques like knowledge": 32648,
      "models pretrained pile": 21847,
      "number training samples": 23171,
      "instruction data quality": 15148,
      "language models substantial": 16717,
      "models llms increasing": 21566,
      "realworld applications understanding": 27335,
      "models trained realworld": 22072,
      "trained realworld dataset": 33422,
      "training data points": 33487,
      "largescale transformerbased language": 17386,
      "models capable handling": 21042,
      "llms recently gained": 18896,
      "summary work contributes": 31634,
      "work contributes improving": 35684,
      "crucial step en": 6448,
      "step en route": 31043,
      "en route enabling": 9280,
      "route enabling widespread": 28990,
      "enabling widespread adoption": 9331,
      "detecting mitigating hallucinations": 7859,
      "like glue superglue": 17864,
      "models downstream tasks": 21173,
      "trained direct preference": 33391,
      "llm large language": 18328,
      "models llms implement": 21560,
      "syntactic semantic information": 31824,
      "increases model size": 14615,
      "models increasingly large": 21391,
      "window size context": 35614,
      "domain natural language": 8576,
      "llama 7b chat": 18068,
      "techniques terms accuracy": 32668,
      "prominent language models": 26267,
      "using single gpu": 34912,
      "matches outperforms stateoftheart": 19655,
      "projection weight matrices": 26253,
      "understanding latent representations": 34244,
      "model sizes existing": 20798,
      "task text generation": 32199,
      "training data quality": 33489,
      "models crucial step": 21117,
      "lowrank adaptation technique": 19302,
      "surge large language": 31726,
      "finetuning specific tasks": 11532,
      "performance language modeling": 24641,
      "empirical evaluations demonstrate": 9223,
      "models gpt4 llama": 21334,
      "arbitrary batch size": 2430,
      "paper introduce simple": 24061,
      "framework experimental results": 11853,
      "language models combinatorial": 16281,
      "solution address challenges": 30469,
      "stateoftheart results natural language": 30985,
      "pretrained transformer language models": 25763,
      "popular pretrained language models": 25138,
      "transformer language models large": 33726,
      "generative pretrained transformer models": 12698,
      "model pretrained language models": 20722,
      "finetuning pretrained language model": 11488,
      "large language models transformer": 17213,
      "instruction finetuned language models": 15156,
      "language models llms necessitates": 16525,
      "sizes 7b 13b 30b": 30295,
      "large language model named": 16983,
      "language models llms increasing": 16504,
      "largescale transformerbased language models": 17387,
      "summary work contributes improving": 31635,
      "crucial step en route": 6449,
      "step en route enabling": 31044,
      "en route enabling widespread": 9281,
      "route enabling widespread adoption": 28991,
      "benchmarks like glue superglue": 3456,
      "language models downstream tasks": 16313,
      "trained direct preference optimization": 33392,
      "llm large language models": 18329,
      "language models llms implement": 16498,
      "language models increasingly large": 16389,
      "domain natural language processing": 8577,
      "paper present novel method": 24093,
      "language models crucial step": 16294,
      "large language models financial": 17022,
      "surge large language models": 31727,
      "language models gpt4 llama": 16374,
      "stateoftheart results natural language processing": 30986,
      "pretrained transformer language models large": 25764,
      "large language models llms necessitates": 17117,
      "large language models llms increasing": 17105,
      "crucial step en route enabling": 6450,
      "step en route enabling widespread": 31045,
      "en route enabling widespread adoption": 9282,
      "llm large language models llms": 18330,
      "large language models llms implement": 17102,
      "domain natural language processing nlp": 8578,
      "surge large language models llms": 31728,
      "large language models gpt4 llama": 17042,
      "goldstandard": 12815,
      "occurrences": 23276,
      "companys": 5071,
      "belong": 3345,
      "court": 6313,
      "fallacy": 10953,
      "genre": 12721,
      "female": 11080,
      "longshort": 19212,
      "15m": 154,
      "amplification": 1893,
      "sparking": 30605,
      "bloomberggpt": 3792,
      "letters": 17728,
      "lawyers": 17421,
      "mediation": 19765,
      "trying": 33948,
      "attracts": 2755,
      "tta": 33951,
      "clearcut": 4648,
      "ecologically": 8806,
      "affairs": 1535,
      "new sampling": 22841,
      "fundamental question": 11981,
      "racial bias": 27144,
      "proposes framework": 26628,
      "ability write": 731,
      "effects observed": 8982,
      "text results": 32933,
      "framework generating": 11862,
      "data challenge": 6630,
      "set augmentation": 29671,
      "set unlabeled": 29713,
      "analysis training": 1973,
      "training nlp": 33574,
      "shown capture": 29871,
      "data unfortunately": 6900,
      "texts providing": 32999,
      "representations neural": 28169,
      "language introduce": 16101,
      "models wild": 22131,
      "generated topic": 12397,
      "language key": 16102,
      "nlp application": 22920,
      "layers predictive": 17444,
      "analysis involves": 1933,
      "way model": 35445,
      "corpus evaluate": 6182,
      "probing framework": 25982,
      "properties experiments": 26471,
      "biases pretrained": 3683,
      "finegrained analysis": 11268,
      "perform human": 24490,
      "test participants": 32777,
      "analyze effect": 1991,
      "plms exhibit": 25044,
      "lexical diversity": 17799,
      "specifically adapted": 30727,
      "present evidence": 25529,
      "showcase models": 29837,
      "accurately identifying": 936,
      "failing meet": 10911,
      "rapid evolution": 27251,
      "models reinforcing": 21917,
      "domains sparking": 8638,
      "sparking great": 30606,
      "unlike proprietary": 34403,
      "unique data": 34358,
      "errors present": 9726,
      "current knowledge": 6495,
      "provides evidence": 26751,
      "method analyzing": 19877,
      "understand potential": 34199,
      "future investigations": 12034,
      "data demonstrated": 6673,
      "analysis critical": 1915,
      "gradientbased methods": 13193,
      "behaviors transformer": 3329,
      "20 large": 210,
      "solving text": 30519,
      "study harness": 31337,
      "emerging field": 9192,
      "gender racial": 12153,
      "computational analysis": 5453,
      "science findings": 29242,
      "ability predict": 711,
      "causal mediation": 4244,
      "automated circuit": 2857,
      "circuit discovery": 4568,
      "datasets addition": 7060,
      "ability acquire": 653,
      "analysis types": 1974,
      "diverse types": 8470,
      "semantics syntax": 29489,
      "predictive results": 25455,
      "features using": 11045,
      "attempt create": 2703,
      "ensuring seamless": 9613,
      "proficiency comprehending": 26179,
      "comprehending generating": 5338,
      "models attributed": 20994,
      "sourced various": 30573,
      "data conducted": 6653,
      "demonstrate reduction": 7487,
      "detection dataset": 7865,
      "light challenges": 17827,
      "reveal inherent": 28800,
      "models tendency": 22051,
      "scarcity largescale": 29193,
      "human judges": 13832,
      "experiments gpt35": 10446,
      "making comprehensive": 19500,
      "humanlike texts": 13911,
      "challenging distinguish": 4382,
      "gpt2 chatgpt": 12877,
      "classify sentiment": 4631,
      "large legal": 17224,
      "ecologically valid": 8807,
      "content control": 5854,
      "approach detect": 2259,
      "using news": 34859,
      "llama increasingly": 18113,
      "techniques large": 32646,
      "using neural language": 34854,
      "gpt2 model way": 12924,
      "paper proposes framework": 24123,
      "gpt2 models trained": 12929,
      "training set augmentation": 33610,
      "set unlabeled data": 29714,
      "training nlp models": 33575,
      "language models wild": 16758,
      "language key challenge": 16103,
      "automatic text generation": 2900,
      "neural network based": 22741,
      "sentiment analysis involves": 29568,
      "based t5 model": 3227,
      "studies shown large": 31284,
      "models code data": 21058,
      "failing meet requirements": 10912,
      "language models bias": 16264,
      "domains sparking great": 8639,
      "unlike proprietary models": 34404,
      "work provides evidence": 35765,
      "conduct quantitative analysis": 5617,
      "sentiment analysis critical": 29567,
      "behaviors transformer models": 3330,
      "llms potential transform": 18859,
      "realworld use case": 27352,
      "automated circuit discovery": 2858,
      "novel method detecting": 23097,
      "gpt4 demonstrated impressive": 13071,
      "proficiency comprehending generating": 26180,
      "language models attributed": 16251,
      "generating humanlike texts": 12430,
      "using neural language models": 34855,
      "neural language models nlms": 22728,
      "models llms potential transform": 21603,
      "large language models attributed": 16993,
      "language models llms potential transform": 16538,
      "effects including": 8980,
      "providing insights": 26775,
      "abilities acquired": 623,
      "text different": 32845,
      "react differently": 27290,
      "performance quickly": 24729,
      "compared transformer": 5179,
      "analysis methods": 1941,
      "output values": 23885,
      "specific subnetworks": 30715,
      "gpt2 gptneo gptj": 12905,
      "deep learning learn": 7330,
      "spanning multiple": 30599,
      "mainly attributed": 19402,
      "explore efficacy": 10583,
      "models latent": 21435,
      "better informed": 3607,
      "introduce multimodal": 15517,
      "tasks 25": 32226,
      "pretrained models latent": 25725,
      "insights generated": 15076,
      "robust foundation": 28932,
      "llms like chatgpt llama": 18758,
      "models llms like chatgpt llama": 21576,
      "framework available": 11831
    }
  }
}