updates

app.py

@@ -10,7 +10,7 @@ from apscheduler.schedulers.background import BackgroundScheduler
 from huggingface_hub import HfApi
 from transformers import AutoConfig
 
-from src.auto_leaderboard.get_model_metadata import apply_metadata
+from src.auto_leaderboard.get_model_metadata import apply_metadata, DO_NOT_SUBMIT_MODELS
 from src.assets.text_content import *
 from src.auto_leaderboard.load_results import get_eval_results_dicts, make_clickable_model
 from src.assets.hardcoded_evals import gpt4_values, gpt35_values, baseline
@@ -82,11 +82,11 @@ def get_leaderboard_df():
         print("Pulling evaluation results for the leaderboard.")
         eval_results_private.git_pull()
 
-    all_data = get_eval_results_dicts(IS_PUBLIC)
+    all_data = get_eval_results_dicts()
 
-    # if not IS_PUBLIC:
-    all_data.append(gpt4_values)
-    all_data.append(gpt35_values)
+    if not IS_PUBLIC:
+        all_data.append(gpt4_values)
+        all_data.append(gpt35_values)
 
     all_data.append(baseline)
     apply_metadata(all_data)  # Populate model type based on known hardcoded values in `metadata.py`
@@ -227,9 +227,13 @@ def add_new_eval(
     os.makedirs(OUT_DIR, exist_ok=True)
     out_path = f"{OUT_DIR}/{model_path}_eval_request_{private}_{precision}_{weight_type}.json"
 
+    # Check if the model has been forbidden:
+    if out_path.split("eval-queue/")[1] in DO_NOT_SUBMIT_MODELS:
+        return styled_warning("Model authors have requested that their model be not submitted on the leaderboard.")
+
     # Check for duplicate submission
     if out_path.split("eval-queue/")[1].lower() in requested_models:
-        return styled_warning("This model has been already submitted.")
+        return styled_warning("This model has been already submitted.")    
 
     with open(out_path, "w") as f:
         f.write(json.dumps(eval_entry))
@@ -290,7 +294,30 @@ def filter_items(df, leaderboard_table, query):
     if AutoEvalColumn.model_type_symbol.name in leaderboard_table.columns:
         filtered_df = df[(df[AutoEvalColumn.model_type_symbol.name] == query)]
     else:
-        return leaderboard_table.columns
+        return filtered_df[leaderboard_table.columns]
+    return filtered_df[leaderboard_table.columns]
+
+def filter_items_size(df, leaderboard_table, query):
+    numeric_intervals = {
+        "all": None,
+        "< 1B": (0, 1),
+        "~3B": (1, 5),
+        "~7B": (6, 11),
+        "~13B": (12, 15),
+        "~35B": (16, 55),
+        "60B+": (55, 1000)
+    }
+
+    if query == "all":
+        return df[leaderboard_table.columns]
+
+    numeric_interval = numeric_intervals[query]
+
+    if AutoEvalColumn.params.name in leaderboard_table.columns:
+        params_column = pd.to_numeric(df[AutoEvalColumn.params.name], errors='coerce')
+        filtered_df = df[params_column.between(*numeric_interval)]
+    else:
+        return filtered_df[leaderboard_table.columns]
     return filtered_df[leaderboard_table.columns]
 
 def change_tab(query_param):
@@ -306,6 +333,10 @@ def change_tab(query_param):
     else:
         return gr.Tabs.update(selected=0)
 
+def update_filter_type(input_type, shown_columns):
+    shown_columns.append(AutoEvalColumn.params.name)
+    return gr.update(visible=(input_type == 'types')), gr.update(visible=(input_type == 'sizes')), shown_columns
+
 
 demo = gr.Blocks(css=custom_css)
 with demo:
@@ -328,20 +359,46 @@ with demo:
                         show_label=False,
                         elem_id="search-bar",
                     )
-                    filter_columns = gr.Radio(
-                        label="⏚ Filter model types",
-                        choices = [
-                            "all", 
-                            ModelType.PT.to_str(),
-                            ModelType.FT.to_str(),
-                            ModelType.IFT.to_str(),
-                            ModelType.RL.to_str(), 
-                        ],
-                        value="all",
-                        elem_id="filter-columns"
-                    )
+                    with gr.Box(elem_id="box-filter"):
+                        filter_type = gr.Dropdown(
+                                label="⏚ Filter model",
+                                choices=["types", "sizes"], value="types",
+                                interactive=True,
+                                elem_id="filter_type"
+                        )
+                        filter_columns = gr.Radio(
+                            label="⏚ Filter model types",
+                            show_label=False,
+                            choices = [
+                                "all", 
+                                ModelType.PT.to_str(),
+                                ModelType.FT.to_str(),
+                                ModelType.IFT.to_str(),
+                                ModelType.RL.to_str(), 
+                            ],
+                            value="all",
+                            elem_id="filter-columns"
+                        )
+                        filter_columns_size = gr.Radio(
+                            label="⏚ Filter model sizes",
+                            show_label=False,
+                            choices = [
+                                "all",
+                                "< 1B",
+                                "~3B",
+                                "~7B",
+                                "~13B",
+                                "~35B",
+                                "60B+"
+                            ],
+                            value="all",
+                            visible=False,
+                            interactive=True,
+                            elem_id="filter-columns-size"
+                        )
+            
             leaderboard_table = gr.components.Dataframe(
-                value=leaderboard_df[[AutoEvalColumn.model_type_symbol.name, AutoEvalColumn.model.name] + shown_columns.value+ [AutoEvalColumn.dummy.name]],
+                value=leaderboard_df[[AutoEvalColumn.model_type_symbol.name, AutoEvalColumn.model.name] + shown_columns.value + [AutoEvalColumn.dummy.name]],
                 headers=[AutoEvalColumn.model_type_symbol.name, AutoEvalColumn.model.name] + shown_columns.value + [AutoEvalColumn.dummy.name],
                 datatype=TYPES,
                 max_rows=None,
@@ -363,8 +420,11 @@ with demo:
                 [hidden_leaderboard_table_for_search, leaderboard_table, search_bar],
                 leaderboard_table,
             )
-            shown_columns.change(select_columns, [hidden_leaderboard_table_for_search, shown_columns], leaderboard_table)
-            filter_columns.change(filter_items, [hidden_leaderboard_table_for_search, leaderboard_table, filter_columns], leaderboard_table)
+            
+            filter_type.change(update_filter_type,inputs=[filter_type, shown_columns],outputs=[filter_columns, filter_columns_size, shown_columns],queue=False).then(select_columns, [hidden_leaderboard_table_for_search, shown_columns], leaderboard_table, queue=False)
+            shown_columns.change(select_columns, [hidden_leaderboard_table_for_search, shown_columns], leaderboard_table, queue=False)
+            filter_columns.change(filter_items, [hidden_leaderboard_table_for_search, leaderboard_table, filter_columns], leaderboard_table, queue=False)
+            filter_columns_size.change(filter_items_size, [hidden_leaderboard_table_for_search, leaderboard_table, filter_columns_size], leaderboard_table, queue=False)
         with gr.TabItem("📝 About", elem_id="llm-benchmark-tab-table", id=2):
             gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")
 
@@ -432,6 +492,6 @@ with demo:
     )
 
 scheduler = BackgroundScheduler()
-scheduler.add_job(restart_space, "interval", seconds=1800)
+scheduler.add_job(restart_space, "interval", seconds=900)
 scheduler.start()
 demo.queue(concurrency_count=40).launch()

src/assets/css_html_js.py

@@ -1,11 +1,4 @@
 custom_css = """
-#changelog-text {
-    font-size: 16px !important;
-}
-
-#changelog-text h2 {
-    font-size: 18px !important;
-}
 
 .markdown-text {
     font-size: 16px !important;
@@ -75,6 +68,38 @@ table th:first-child {
 #scale-logo .download {
     display: none;
 }
+#filter_type{
+    border: 0;
+    padding-left: 0;
+    padding-top: 0;
+}
+#filter_type label {
+    display: flex;
+}
+#filter_type label > span{
+    margin-top: var(--spacing-lg);
+    margin-right: 0.5em;
+}
+#filter_type label > .wrap{
+    width: 103px;
+}
+#filter_type label > .wrap .wrap-inner{  
+    padding: 2px;
+}
+#filter_type label > .wrap .wrap-inner input{
+    width: 1px
+}
+#filter-columns{
+    border:0;
+    padding:0;
+}
+#filter-columns-size{
+    border:0;
+    padding:0;
+}
+#box-filter > .form{
+    border: 0
+}
 """
 
 get_window_url_params = """
@@ -83,4 +108,4 @@ get_window_url_params = """
         url_params = Object.fromEntries(params);
         return url_params;
     }
-    """
+    """

src/assets/text_content.py

@@ -1,76 +1,31 @@
 from ..auto_leaderboard.model_metadata_type import ModelType
 
-CHANGELOG_TEXT = f"""
-## [2023-06-19]
-- Added model type column
-- Hid revision and 8bit columns since all models are the same atm
-
-## [2023-06-16]
-- Refactored code base
-- Added new columns: number of parameters, hub likes, license
-
-## [2023-06-13] 
-- Adjust description for TruthfulQA
-
-## [2023-06-12] 
-- Add Human & GPT-4 Evaluations
-
-## [2023-06-05] 
-- Increase concurrent thread count to 40
-- Search models on ENTER
-
-## [2023-06-02] 
-- Add a typeahead search bar
-- Use webhooks to automatically spawn a new Space when someone opens a PR
-- Start recording `submitted_time` for eval requests
-- Limit AutoEvalColumn max-width
-
-## [2023-05-30] 
-- Add a citation button
-- Simplify Gradio layout
-
-## [2023-05-29] 
-- Auto-restart every hour for the latest results
-- Sync with the internal version (minor style changes)
-
-## [2023-05-24] 
-- Add a baseline that has 25.0 for all values
-- Add CHANGELOG
-
-## [2023-05-23] 
-- Fix a CSS issue that made the leaderboard hard to read in dark mode
-
-## [2023-05-22] 
-- Display a success/error message after submitting evaluation requests
-- Reject duplicate submission
-- Do not display results that have incomplete results 
-- Display different queues for jobs that are RUNNING, PENDING, FINISHED status 
-
-## [2023-05-15] 
-- Fix a typo: from "TruthQA" to "QA"
-
-## [2023-05-10] 
-- Fix a bug that prevented auto-refresh
-
-## [2023-05-10] 
-- Release the leaderboard to public
-"""
-
 TITLE = """<h1 align="center" id="space-title">🤗 Open LLM Leaderboard</h1>"""
 
 INTRODUCTION_TEXT = f"""
-📐 The 🤗 Open LLM Leaderboard aims to track, rank and evaluate LLMs and chatbots as they are released. 
-
-🤗 Anyone from the community can submit a model for automated evaluation on the 🤗 GPU cluster, as long as it is a 🤗 Transformers model with weights on the Hub. We also support evaluation of models with delta-weights for non-commercial licensed models, such as the original LLaMa release.
+📐 The 🤗 Open LLM Leaderboard aims to track, rank and evaluate open LLMs and chatbots.
 
-Other cool benchmarks for LLMs are developed at HuggingFace, go check them out: 🙋🤖 [human and GPT4 evals](https://huggingface.co/spaces/HuggingFaceH4/human_eval_llm_leaderboard), 🖥️ [performance benchmarks](https://huggingface.co/spaces/optimum/llm-perf-leaderboard)
+🤗 Submit a model for automated evaluation on the 🤗 GPU cluster on the "Submit" page! 
+The leaderboard's backend runs the great [Eleuther AI Language Model Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness) - read more details in the "About" page!
 """
 
 LLM_BENCHMARKS_TEXT = f"""
 # Context
 With the plethora of large language models (LLMs) and chatbots being released week upon week, often with grandiose claims of their performance, it can be hard to filter out the genuine progress that is being made by the open-source community and which model is the current state of the art. 
 
-📈 We evaluate models on 4 key benchmarks from the <a href="https://github.com/EleutherAI/lm-evaluation-harness" target="_blank">  Eleuther AI Language Model Evaluation Harness </a>, a unified framework to test generative language models on a large number of different evaluation tasks. 
+## Icons
+{ModelType.PT.to_str(" : ")} model
+{ModelType.FT.to_str(" : ")} model
+{ModelType.IFT.to_str(" : ")} model
+{ModelType.RL.to_str(" : ")} model
+If there is no icon, we have not uploaded the information on the model yet, feel free to open an issue with the model information!
+
+🏴‍☠️ indicates that this model has been flagged by the community, and should probably be ignored! Clicking the icon will redirect you to the discussion about the model.
+(For ex, the model was trained on the evaluation data, and is therefore cheating on the leaderboard.)
+
+## How it works
+
+📈 We evaluate models on 4 key benchmarks using the <a href="https://github.com/EleutherAI/lm-evaluation-harness" target="_blank">  Eleuther AI Language Model Evaluation Harness </a>, a unified framework to test generative language models on a large number of different evaluation tasks. 
 
 - <a href="https://arxiv.org/abs/1803.05457" target="_blank">  AI2 Reasoning Challenge </a> (25-shot) - a set of grade-school science questions.
 - <a href="https://arxiv.org/abs/1905.07830" target="_blank">  HellaSwag </a> (10-shot) - a test of commonsense inference, which is easy for humans (~95%) but challenging for SOTA models.
@@ -80,38 +35,13 @@ With the plethora of large language models (LLMs) and chatbots being released we
 For all these evaluations, a higher score is a better score. 
 We chose these benchmarks as they test a variety of reasoning and general knowledge across a wide variety of fields in 0-shot and few-shot settings.
 
-# Some good practices before submitting a model
-
-### 1) Make sure you can load your model and tokenizer using AutoClasses:
-```python
-from transformers import AutoConfig, AutoModel, AutoTokenizer
-config = AutoConfig.from_pretrained("your model name", revision=revision)
-model = AutoModel.from_pretrained("your model name", revision=revision)
-tokenizer = AutoTokenizer.from_pretrained("your model name", revision=revision)
-```
-If this step fails, follow the error messages to debug your model before submitting it. It's likely your model has been improperly uploaded. 
-
-Note: make sure your model is public!
-Note: if your model needs `use_remote_code=True`, we do not support this option yet but we are working on adding it, stay posted!
-
-### 2) Convert your model weights to [safetensors](https://huggingface.co/docs/safetensors/index)
-It's a new format for storing weights which is safer and faster to load and use. It will also allow us to add the number of parameters of your model to the `Extended Viewer`!
-
-### 3) Make sure your model has an open license!
-This is a leaderboard for Open LLMs, and we'd love for as many people as possible to know they can use your model 🤗
-
-### 4) Fill up your model card
-When we add extra information about models to the leaderboard, it will be automatically taken from the model card
-
-# Reproducibility and details
-
-### Details and logs
+## Details and logs
 You can find:
 - detailed numerical results in the `results` Hugging Face dataset: https://huggingface.co/datasets/open-llm-leaderboard/results
 - details on the input/outputs for the models in the `details` Hugging Face dataset: https://huggingface.co/datasets/open-llm-leaderboard/details
 - community queries and running status in the `requests` Hugging Face dataset: https://huggingface.co/datasets/open-llm-leaderboard/requests
 
-### Reproducibility
+## Reproducibility
 To reproduce our results, here is the commands you can run, using [this version](https://github.com/EleutherAI/lm-evaluation-harness/tree/b281b0921b636bc36ad05c0b0b0763bd6dd43463) of the Eleuther AI Harness:
 `python main.py --model=hf-causal --model_args="pretrained=<your_model>,use_accelerate=True,revision=<your_model_revision>"`
 ` --tasks=<task_list> --num_fewshot=<n_few_shot> --batch_size=2 --output_path=<output_path>`
@@ -125,29 +55,48 @@ The tasks and few shots parameters are:
 - TruthfulQA: 0-shot, *truthfulqa-mc* (`mc2`)
 - MMLU: 5-shot, *hendrycksTest-abstract_algebra,hendrycksTest-anatomy,hendrycksTest-astronomy,hendrycksTest-business_ethics,hendrycksTest-clinical_knowledge,hendrycksTest-college_biology,hendrycksTest-college_chemistry,hendrycksTest-college_computer_science,hendrycksTest-college_mathematics,hendrycksTest-college_medicine,hendrycksTest-college_physics,hendrycksTest-computer_security,hendrycksTest-conceptual_physics,hendrycksTest-econometrics,hendrycksTest-electrical_engineering,hendrycksTest-elementary_mathematics,hendrycksTest-formal_logic,hendrycksTest-global_facts,hendrycksTest-high_school_biology,hendrycksTest-high_school_chemistry,hendrycksTest-high_school_computer_science,hendrycksTest-high_school_european_history,hendrycksTest-high_school_geography,hendrycksTest-high_school_government_and_politics,hendrycksTest-high_school_macroeconomics,hendrycksTest-high_school_mathematics,hendrycksTest-high_school_microeconomics,hendrycksTest-high_school_physics,hendrycksTest-high_school_psychology,hendrycksTest-high_school_statistics,hendrycksTest-high_school_us_history,hendrycksTest-high_school_world_history,hendrycksTest-human_aging,hendrycksTest-human_sexuality,hendrycksTest-international_law,hendrycksTest-jurisprudence,hendrycksTest-logical_fallacies,hendrycksTest-machine_learning,hendrycksTest-management,hendrycksTest-marketing,hendrycksTest-medical_genetics,hendrycksTest-miscellaneous,hendrycksTest-moral_disputes,hendrycksTest-moral_scenarios,hendrycksTest-nutrition,hendrycksTest-philosophy,hendrycksTest-prehistory,hendrycksTest-professional_accounting,hendrycksTest-professional_law,hendrycksTest-professional_medicine,hendrycksTest-professional_psychology,hendrycksTest-public_relations,hendrycksTest-security_studies,hendrycksTest-sociology,hendrycksTest-us_foreign_policy,hendrycksTest-virology,hendrycksTest-world_religions* (average of all the results `acc`)
 
-### Quantization
+## Quantization
 To get more information about quantization, see:
 - 8 bits: [blog post](https://huggingface.co/blog/hf-bitsandbytes-integration), [paper](https://arxiv.org/abs/2208.07339)
 - 4 bits: [blog post](https://huggingface.co/blog/4bit-transformers-bitsandbytes), [paper](https://arxiv.org/abs/2305.14314)
 
-### Icons
-{ModelType.PT.to_str(" : ")} model
-{ModelType.FT.to_str(" : ")} model
-{ModelType.IFT.to_str(" : ")} model
-{ModelType.RL.to_str(" : ")} model
-If there is no icon, we have not uploaded the information on the model yet, feel free to open an issue with the model information!
+## More resources
+If you still have questions, you can check our FAQ [here](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/179)!
+We also gather cool resources from the community, other teams, and other labs [here](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/174)!
+"""
 
+EVALUATION_QUEUE_TEXT = f"""
+# Evaluation Queue for the ��� Open LLM Leaderboard
+
+Models added here will be automatically evaluated on the 🤗 cluster.
 
-# In case of model failure
+## Some good practices before submitting a model
+
+### 1) Make sure you can load your model and tokenizer using AutoClasses:
+```python
+from transformers import AutoConfig, AutoModel, AutoTokenizer
+config = AutoConfig.from_pretrained("your model name", revision=revision)
+model = AutoModel.from_pretrained("your model name", revision=revision)
+tokenizer = AutoTokenizer.from_pretrained("your model name", revision=revision)
+```
+If this step fails, follow the error messages to debug your model before submitting it. It's likely your model has been improperly uploaded. 
+
+Note: make sure your model is public!
+Note: if your model needs `use_remote_code=True`, we do not support this option yet but we are working on adding it, stay posted!
+
+### 2) Convert your model weights to [safetensors](https://huggingface.co/docs/safetensors/index)
+It's a new format for storing weights which is safer and faster to load and use. It will also allow us to add the number of parameters of your model to the `Extended Viewer`!
+
+### 3) Make sure your model has an open license!
+This is a leaderboard for Open LLMs, and we'd love for as many people as possible to know they can use your model 🤗
+
+### 4) Fill up your model card
+When we add extra information about models to the leaderboard, it will be automatically taken from the model card
+
+## In case of model failure
 If your model is displayed in the `FAILED` category, its execution stopped. 
 Make sure you have followed the above steps first. 
 If everything is done, check you can launch the EleutherAIHarness on your model locally, using the above command without modifications (you can add `--limit` to limit the number of examples per task).
-
-"""
-
-EVALUATION_QUEUE_TEXT = f"""
-# Evaluation Queue for the 🤗 Open LLM Leaderboard
-These models will be automatically evaluated on the 🤗 cluster.
 """
 
 CITATION_BUTTON_LABEL = "Copy the following snippet to cite these results"
@@ -216,4 +165,4 @@ CITATION_BUTTON_TEXT = r"""
       eprint={2109.07958},
       archivePrefix={arXiv},
       primaryClass={cs.CL}
-}"""
+}"""

src/auto_leaderboard/get_model_metadata.py

@@ -1,9 +1,14 @@
 import re
 import os
+import glob
+import json
+import os
 from typing import List
+from tqdm import tqdm
 
-from src.utils_display import AutoEvalColumn
-from src.auto_leaderboard.model_metadata_type import get_model_type
+from src.utils_display import AutoEvalColumn, model_hyperlink
+from src.auto_leaderboard.model_metadata_type import ModelType, model_type_from_str, MODEL_TYPE_METADATA
+from src.auto_leaderboard.model_metadata_flags import FLAGGED_MODELS, DO_NOT_SUBMIT_MODELS
 
 from huggingface_hub import HfApi
 import huggingface_hub
@@ -11,7 +16,7 @@ api = HfApi(token=os.environ.get("H4_TOKEN", None))
 
 
 def get_model_infos_from_hub(leaderboard_data: List[dict]):
-    for model_data in leaderboard_data:
+    for model_data in tqdm(leaderboard_data):
         model_name = model_data["model_name_for_query"]
         try:
             model_info = api.model_info(model_name)
@@ -21,12 +26,6 @@ def get_model_infos_from_hub(leaderboard_data: List[dict]):
             model_data[AutoEvalColumn.likes.name] = None
             model_data[AutoEvalColumn.params.name] = get_model_size(model_name, None)
             continue
-        except Exception:
-            print("Repo fetch error!", model_name)
-            model_data[AutoEvalColumn.license.name] = None
-            model_data[AutoEvalColumn.likes.name] = None
-            model_data[AutoEvalColumn.params.name] = None
-            continue
 
         model_data[AutoEvalColumn.license.name] = get_model_license(model_info)
         model_data[AutoEvalColumn.likes.name] = get_model_likes(model_info)
@@ -57,6 +56,68 @@ def get_model_size(model_name, model_info):
             return None
 
 
+def get_model_type(leaderboard_data: List[dict]):
+    for model_data in leaderboard_data:
+        request_files = os.path.join("eval-queue", model_data["model_name_for_query"] + "_eval_request_*" + ".json")
+        request_files = glob.glob(request_files)
+
+        # Select correct request file (precision)
+        request_file = ""
+        if len(request_files) == 1:
+            request_file = request_files[0]
+        elif len(request_files) > 1:
+            request_files = sorted(request_files, reverse=True)
+            for tmp_request_file in request_files:
+                with open(tmp_request_file, "r") as f:
+                    req_content = json.load(f)
+                    if req_content["status"] == "FINISHED" and req_content["precision"] == model_data["Precision"].split(".")[-1]: 
+                        request_file = tmp_request_file
+        
+        if request_file == "":
+            model_data[AutoEvalColumn.model_type.name] = ""
+            model_data[AutoEvalColumn.model_type_symbol.name] = ""
+            continue
+
+        try:
+            with open(request_file, "r") as f:
+                request = json.load(f)
+            is_delta = request["weight_type"] != "Original"
+        except Exception:
+            is_delta = False
+
+        try:
+            with open(request_file, "r") as f:
+                request = json.load(f)
+            model_type = model_type_from_str(request["model_type"])
+            model_data[AutoEvalColumn.model_type.name] = model_type.value.name
+            model_data[AutoEvalColumn.model_type_symbol.name] = model_type.value.symbol #+ ("🔺" if is_delta else "")
+        except KeyError:
+            if model_data["model_name_for_query"] in MODEL_TYPE_METADATA:
+                model_data[AutoEvalColumn.model_type.name] = MODEL_TYPE_METADATA[model_data["model_name_for_query"]].value.name
+                model_data[AutoEvalColumn.model_type_symbol.name] = MODEL_TYPE_METADATA[model_data["model_name_for_query"]].value.symbol #+ ("🔺" if is_delta else "")
+            else:
+                model_data[AutoEvalColumn.model_type.name] = ModelType.Unknown.value.name
+                model_data[AutoEvalColumn.model_type_symbol.name] = ModelType.Unknown.value.symbol
+
+def flag_models(leaderboard_data:List[dict]):
+    for model_data in leaderboard_data:
+        if model_data["model_name_for_query"] in FLAGGED_MODELS:
+            issue_num = FLAGGED_MODELS[model_data["model_name_for_query"]].split("/")[-1]
+            issue_link = model_hyperlink(FLAGGED_MODELS[model_data["model_name_for_query"]], f"See discussion #{issue_num}")
+            model_data[AutoEvalColumn.model.name] =  f"{model_data[AutoEvalColumn.model.name]} has been flagged! {issue_link}"
+
+def remove_forbidden_models(leaderboard_data: List[dict]):
+    indices_to_remove = []
+    for ix, model in enumerate(leaderboard_data):
+        if model["model_name_for_query"] in DO_NOT_SUBMIT_MODELS:
+            indices_to_remove.append(ix)
+
+    for ix in reversed(indices_to_remove):
+        leaderboard_data.pop(ix)
+    return leaderboard_data
+
 def apply_metadata(leaderboard_data: List[dict]):
+    leaderboard_data = remove_forbidden_models(leaderboard_data)
     get_model_type(leaderboard_data)
     get_model_infos_from_hub(leaderboard_data)
+    flag_models(leaderboard_data)

src/auto_leaderboard/load_results.py

@@ -91,7 +91,7 @@ def parse_eval_result(json_filepath: str) -> Tuple[str, list[dict]]:
 
     eval_results = []
     for benchmark, metric in zip(BENCHMARKS, METRICS):
-        accs = np.array([v[metric] for k, v in data["results"].items() if benchmark in k])
+        accs = np.array([v.get(metric, 0) for k, v in data["results"].items() if benchmark in k])
         if accs.size == 0:
             continue
         mean_acc = np.mean(accs) * 100.0
@@ -102,7 +102,7 @@ def parse_eval_result(json_filepath: str) -> Tuple[str, list[dict]]:
     return result_key, eval_results
 
 
-def get_eval_results(is_public) -> List[EvalResult]:
+def get_eval_results() -> List[EvalResult]:
     json_filepaths = []
 
     for root, dir, files in os.walk("eval-results"):
@@ -135,7 +135,7 @@ def get_eval_results(is_public) -> List[EvalResult]:
     return eval_results
 
 
-def get_eval_results_dicts(is_public=True) -> List[Dict]:
-    eval_results = get_eval_results(is_public)
+def get_eval_results_dicts() -> List[Dict]:
+    eval_results = get_eval_results()
 
     return [e.to_dict() for e in eval_results]

src/auto_leaderboard/model_metadata_flags.py

@@ -0,0 +1,12 @@
+# Models which have been flagged by users as being problematic for a reason or another
+# (Model name to forum discussion link)
+FLAGGED_MODELS = {
+    "Voicelab/trurl-2-13b": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/202",
+    "deepnight-research/llama-2-70B-inst": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/207",
+    "Aspik101/trurl-2-13b-pl-instruct_unload": "https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/discussions/213",
+}
+
+# Models which have been requested by orgs to not be submitted on the leaderboard
+DO_NOT_SUBMIT_MODELS = [
+    "Voicelab/trurl-2-13b", # trained on MMLU
+]

src/auto_leaderboard/model_metadata_type.py

@@ -1,35 +1,26 @@
 from dataclasses import dataclass
 from enum import Enum
-import glob
-import json
-import os
-from typing import Dict, List
+from typing import Dict
 
-from ..utils_display import AutoEvalColumn
 
 @dataclass
 class ModelInfo:
     name: str
     symbol: str # emoji
 
-model_type_symbols = {
-    "fine-tuned": "🔶",
-    "pretrained": "🟢",
-    "RL-tuned": "🟦",
-    "instruction-tuned": "⭕",
-}
 
 class ModelType(Enum):
     PT = ModelInfo(name="pretrained", symbol="🟢")
     FT = ModelInfo(name="fine-tuned", symbol="🔶")
     IFT = ModelInfo(name="instruction-tuned", symbol="⭕")
     RL = ModelInfo(name="RL-tuned", symbol="🟦")
+    Unknown = ModelInfo(name="Unknown, add type to request file!", symbol="?")
 
     def to_str(self, separator = " "):
         return f"{self.value.symbol}{separator}{self.value.name}" 
 
 
-TYPE_METADATA: Dict[str, ModelType] = {
+MODEL_TYPE_METADATA: Dict[str, ModelType] = {
     'notstoic/PygmalionCoT-7b': ModelType.IFT,
     'aisquared/dlite-v1-355m': ModelType.IFT,
     'aisquared/dlite-v1-1_5b': ModelType.IFT,
@@ -211,10 +202,10 @@ TYPE_METADATA: Dict[str, ModelType] = {
     'dvruette/oasst-gpt-neox-20b-1000-steps': ModelType.FT,
     'dvruette/llama-13b-pretrained-dropout': ModelType.PT,
     'dvruette/llama-13b-pretrained': ModelType.PT,
-    'dvruette/llama-13b-pretrained-sft-epoch-1': ModelType.PT,
-    'dvruette/llama-13b-pretrained-sft-do2': ModelType.PT,
+    'dvruette/llama-13b-pretrained-sft-epoch-1': ModelType.FT,
+    'dvruette/llama-13b-pretrained-sft-do2': ModelType.FT,
     'dvruette/oasst-gpt-neox-20b-3000-steps': ModelType.FT,
-    'dvruette/oasst-pythia-12b-pretrained-sft': ModelType.PT,
+    'dvruette/oasst-pythia-12b-pretrained-sft': ModelType.FT,
     'dvruette/oasst-pythia-6.9b-4000-steps': ModelType.FT,
     'dvruette/gpt-neox-20b-full-precision': ModelType.FT,
     'dvruette/oasst-llama-13b-1000-steps': ModelType.FT,
@@ -547,33 +538,14 @@ TYPE_METADATA: Dict[str, ModelType] = {
 }
 
 
-def get_model_type(leaderboard_data: List[dict]):
-    for model_data in leaderboard_data:
-        # Todo @clefourrier once requests are connected with results 
-        # Stored information
-        request_file = os.path.join("eval-queue", model_data["model_name_for_query"] + "_eval_request_*" + ".json")
-        request_file = glob.glob(request_file)
-
-        if len(request_file) == 0:
-            model_data[AutoEvalColumn.model_type.name] = ""
-            model_data[AutoEvalColumn.model_type_symbol.name] = ""
-            continue
-
-        request_file = request_file[0]
-
-        try:
-            with open(request_file, "r") as f:
-                request = json.load(f)
-            is_delta = request["weight_type"] != "Original"
-        except Exception:
-            is_delta = False
+def model_type_from_str(type):
+    if "fine-tuned" in type or "🔶" in type:
+        return ModelType.FT
+    if "pretrained" in type or "🟢" in type:
+        return ModelType.PT
+    if "RL-tuned" in type or "🟦" in type:
+        return ModelType.RL
+    if "instruction-tuned" in type or "⭕" in type:
+        return ModelType.IFT
+    return ModelType.Unknown
 
-        try:
-            with open(request_file, "r") as f:
-                request = json.load(f)
-            model_type = request["model_type"]
-            model_data[AutoEvalColumn.model_type.name] = model_type
-            model_data[AutoEvalColumn.model_type_symbol.name] = model_type_symbols[model_type] + ("🔺" if is_delta else "")
-        except Exception:
-            model_data[AutoEvalColumn.model_type.name] = "Unknown, add type to request file!"
-            model_data[AutoEvalColumn.model_type_symbol.name] = "?"

src/init.py

@@ -1,8 +1,6 @@
 import os 
 from huggingface_hub import Repository
 
-H4_TOKEN = os.environ.get("H4_TOKEN", None)
-
 
 def get_all_requested_models(requested_models_dir):
     depth = 1
@@ -20,28 +18,23 @@ def load_all_info_from_hub(QUEUE_REPO, RESULTS_REPO, QUEUE_PATH, RESULTS_PATH):
     eval_results_repo = None
     requested_models = None
 
-    if H4_TOKEN:
-        print("Pulling evaluation requests and results.")
-
-        eval_queue_repo = Repository(
-            local_dir=QUEUE_PATH,
-            clone_from=QUEUE_REPO,
-            use_auth_token=H4_TOKEN,
-            repo_type="dataset",
-        )
-        eval_queue_repo.git_pull()
-
-        eval_results_repo = Repository(
-            local_dir=RESULTS_PATH,
-            clone_from=RESULTS_REPO,
-            use_auth_token=H4_TOKEN,
-            repo_type="dataset",
-        )
-        eval_results_repo.git_pull()
-
-        requested_models = get_all_requested_models("eval-queue")
-    else:
-        print("No HuggingFace token provided. Skipping evaluation requests and results.")
+    print("Pulling evaluation requests and results.")
+
+    eval_queue_repo = Repository(
+        local_dir=QUEUE_PATH,
+        clone_from=QUEUE_REPO,
+        repo_type="dataset",
+    )
+    eval_queue_repo.git_pull()
+
+    eval_results_repo = Repository(
+        local_dir=RESULTS_PATH,
+        clone_from=RESULTS_REPO,
+        repo_type="dataset",
+    )
+    eval_results_repo.git_pull()
+
+    requested_models = get_all_requested_models("eval-queue")
 
     return eval_queue_repo, requested_models, eval_results_repo
 

src/utils_display.py

@@ -1,4 +1,9 @@
+import os
 from dataclasses import dataclass
+from huggingface_hub import HfApi
+
+API = HfApi()
+
 
 # These classes are for user facing column names, to avoid having to change them 
 # all around the code when a modif is needed 
@@ -84,10 +89,24 @@ def make_clickable_model(model_name):
         link = KOALA_LINK
     elif model_name == "oasst-12b":
         link = OASST_LINK
-    #else:
-    #    link = MODEL_PAGE
-  
-    return model_hyperlink(link, model_name)
+
+    details_model_name = model_name.replace('/', '__')
+    details_link = f"https://huggingface.co/datasets/open-llm-leaderboard/details_{details_model_name}"
+
+    if not bool(os.getenv("DEBUG", "False")):
+        # We only add these checks when not debugging, as they are extremely slow
+        print(f"details_link: {details_link}")
+        try:
+            check_path = list(API.list_files_info(repo_id=f"open-llm-leaderboard/details_{details_model_name}",
+                                                paths="README.md",
+                                                repo_type="dataset"))
+            print(f"check_path: {check_path}")
+        except Exception as err:
+            # No details repo for this model
+            print(f"No details repo for this model: {err}")
+            return model_hyperlink(link, model_name)
+
+    return model_hyperlink(link, model_name) + '  ' + model_hyperlink(details_link, "📑")
 
 def styled_error(error):
     return f"<p style='color: red; font-size: 20px; text-align: center;'>{error}</p>"