I briefly reviewed the paper "SFT Memorizes, RL Generalizes," which compares SFT and RL in post-training of LLM/VLM from HKU, UC Berkeley, Google DeepMind, and New York University
The conclusion suggests SFT excels in memorization, while RL is better for generalization. However, since LLM/VLM should benefit humans beyond just generalization, a mix of SFT and RL is advisable. Typically, some SFT is followed by RL to understand prompt formats and enhance generalization through trial and error.
The study focused on one model, Llama-3.2-Vision-11B, using environments like General Points for arithmetic reasoning and V-IRL for spatial reasoning. Training data was used for both SFT and RL, with evaluations on in-distribution and out-of-distribution data to assess memorization and generalization.
I want to apply RL extensively, but it requires building a similar simulation environment. For domain-specific models, significant investment in creating a "playground" for the model is crucial, as the effort will directly influence the outcomes.
The OpenAI o3-mini model is a significant improvement over the o1-mini, reaching o1 performance levels. While generally good, its performance isn't universally better than previous models (o1, o1-prev.) or GPT-4o across all benchmarks. This means workflows should be re-evaluated with each model upgrade.
The o3-mini has "low," "medium," and "high" versions, with "low" being the base model used for benchmarking. It's speculated that the higher versions simply involve more processing. A fair comparison with other models like Gemini 2.0 Thinking or DeepSeek-R1 would likely need to use the "low" version and a similar "think more" mechanism.
The system card is recommended reading due to its comprehensive benchmark data.
This week in open AI was 🔥 Let's recap! 🤗 merve/january-31-releases-679a10669bd4030090c5de4d LLMs 💬 > Huge: AllenAI released new Tülu models that outperform DeepSeek R1 using Reinforcement Learning with Verifiable Reward (RLVR) based on Llama 3.1 405B 🔥 > Mistral AI is back to open-source with their "small" 24B models (base & SFT), with Apache 2.0 license 😱 > Alibaba Qwen released their 1M context length models Qwen2.5-Instruct-1M, great for agentic use with Apache 2.0 license 🔥 > Arcee AI released Virtuoso-medium, 32.8B LLMs distilled from DeepSeek V3 with dataset of 5B+ tokens > Velvet-14B is a new family of 14B Italian LLMs trained on 10T tokens in six languages > OpenThinker-7B is fine-tuned version of Qwen2.5-7B-Instruct on OpenThoughts dataset
VLMs & vision 👀 > Alibaba Qwen is back with Qwen2.5VL, amazing new capabilities ranging from agentic computer use to zero-shot localization 🔥 > NVIDIA released new series of Eagle2 models with 1B and 9B sizes > DeepSeek released Janus-Pro, new any-to-any model (image-text generation from image-text input) with MIT license > BEN2 is a new background removal model with MIT license!
Audio 🗣️ > YuE is a new open-source music generation foundation model, lyrics-to-song generation
Explain like i'm 5 the last take from @thomwolf on X about Dario's essay on DeepSeek:
—› Open-source AI is like a big cookbook that everyone can read and improve. Instead of a few chefs keeping their recipes secret, anyone can cook, test, and invent new things.
If only one company controls AI, everything stops if they have a problem—like when the internet goes down. With open-source, many people can help, making sure it keeps running smoothly.
AI isn’t just a race between two countries; it’s a team effort around the world. By sharing, we move faster and create safer technology for everyone. — 🤗
Simple summary on DeepSeek AI's Janus-Pro: A fresh take on multimodal AI! It builds on its predecessor, Janus, by tweaking the training methodology rather than the model architecture. The result? Improved performance in understanding and generating multimodal data.
Janus-Pro uses a three-stage training strategy, similar to Janus, but with key modifications: ✦ Stage 1 & 2: Focus on separate training for specific objectives, rather than mixing data. ✦ Stage 3: Fine-tuning with a careful balance of multimodal data.
Benchmarks show Janus-Pro holds its own against specialized models like TokenFlow XL and MetaMorph, and other multimodal models like SD3 Medium and DALL-E 3.
The main limitation? Low image resolution (384x384). However, this seems like a strategic choice to focus on establishing a solid "recipe" for multimodal models. Future work will likely leverage this recipe and increased computing power to achieve higher resolutions.
Multimodal 💬 - We have released SmolVLM -- tiniest VLMs that come in 256M and 500M, with it's retrieval models ColSmol for multimodal RAG 💗 - UI-TARS are new models by ByteDance to unlock agentic GUI control 🤯 in 2B, 7B and 72B - Alibaba DAMO lab released VideoLlama3, new video LMs that come in 2B and 7B - MiniMaxAI released Minimax-VL-01, where decoder is based on MiniMax-Text-01 456B MoE model with long context - Dataset: Yale released a new benchmark called MMVU - Dataset: CAIS released Humanity's Last Exam (HLE) a new challenging MM benchmark
LLMs 📖 - DeepSeek-R1 & DeepSeek-R1-Zero: gigantic 660B reasoning models by DeepSeek, and six distilled dense models, on par with o1 with MIT license! 🤯 - Qwen2.5-Math-PRM: new math models by Qwen in 7B and 72B - NVIDIA released AceMath and AceInstruct, new family of models and their datasets (SFT and reward ones too!)
Audio 🗣️ - Llasa is a new speech synthesis model based on Llama that comes in 1B,3B, and 8B - TangoFlux is a new audio generation model trained from scratch and aligned with CRPO
Image/Video/3D Generation ⏯️ - Flex.1-alpha is a new 8B pre-trained diffusion model by ostris similar to Flux - tencent released Hunyuan3D-2, new 3D asset generation from images
smolagents can see 🔥 we just shipped vision support to smolagents 🤗 agentic computers FTW
you can now: 💻 let the agent get images dynamically (e.g. agentic web browser) 📑 pass images at the init of the agent (e.g. chatting with documents, filling forms automatically etc) with few LoC change! 🤯 you can use transformers models locally (like Qwen2VL) OR plug-in your favorite multimodal inference provider (gpt-4o, antrophic & co) 🤠
New look for AI powered paper reviews from the list by Hugging Face Daily Papers ( managed by the @akhaliq )
Bookmark the webpage along, check comprehensive reviews by Google DeepMind Gemini 1.5, and listen to audio podcast made by the same tech used in NotebookLM. Link: https://deep-diver.github.io/ai-paper-reviewer/
This is not an official service by Hugging Face. It is just a service developed by an individual developer using his own money :)
I just released Sentence Transformers v3.4.0, featuring a memory leak fix, compatibility between the powerful Cached... losses and the Matryoshka loss modifier, and a bunch of fixes & small features.
🪆 Matryoshka & Cached loss compatibility It is now possible to combine the powerful Cached... losses (which use in-batch negatives & a caching mechanism to allow for endless batch size & negatives) with the Matryoshka loss modifier which modifies a base loss such that it is trained not only on the maximum dimensionality (e.g. 1024 dimensions), but also on many lower dimensions (e.g. 768, 512, 256, 128, 64, 32). After training, these models' embeddings can be truncated for faster retrieval, etc.
🎞️ Resolve memory leak when Model and Trainer are reinitialized Due to a circular dependency between Trainer -> Model -> ModelCardData -> Trainer, deleting both the trainer & model still didn't free up the memory. This led to a memory leak in scripts where you repeatedly do so.
➕ New Features Many new small features, e.g. multi-GPU support for 'mine_hard_negatives', a 'margin' parameter to TripletEvaluator, and Matthews Correlation Coefficient in the BinaryClassificationEvaluator.
🐛 Bug Fixes Also a bunch of fixes, for example that subsequent batches were not sorted when using the "no_duplicates" batch sampler. See the release notes for more details.
Simple summarization of Evolving Deeper LLM Thinking (Google DeepMind)
The process starts by posing a question. 1) The LLM generates initial responses. 2) These generated responses are evaluated according to specific criteria (program-based checker). 3) The LLM critiques the evaluated results. 4) The LLM refines the responses based on the evaluation, critique, and original responses.
The refined response is then fed back into step 2). If it meets the criteria, the process ends. Otherwise, the algorithm generates more responses based on the refined ones (with some being discarded, some remaining, and some responses potentially being merged).
Through this process, it demonstrated excellent performance in complex scheduling problems (travel planning, meeting scheduling, etc.). It's a viable method for finding highly effective solutions in specific scenarios.
However, there are two major drawbacks: 🤔 An excessive number of API calls are required. (While the cost might not be very high, it leads to significant latency.) 🤔 The evaluator is program-based. (This limits its use as a general method. It could potentially be modified/implemented using LLM as Judge, but that would introduce additional API costs for evaluation.)
Simple Summarization on DeepSeek-R1 from DeepSeek AI
The RL stage is very important. ↳ However, it is difficult to create a truly helpful AI for people solely through RL. ↳ So, we applied a learning pipeline consisting of four stages: providing a good starting point, reasoning RL, SFT, and safety RL, and achieved performance comparable to o1. ↳ Simply fine-tuning other open models with the data generated by R1-Zero (distillation) resulted in performance comparable to o1-mini.
Of course, this is just a brief overview and may not be of much help. All models are accessible on Hugging Face, and the paper can be read through the GitHub repository.
👀 Multimodal - MiniCPM-o 2.6 is a new sota any-to-any model by OpenBMB (vision, speech and text!) - VideoChat-Flash-Qwen2.5-2B is new video multimodal models by OpenGVLab that come in sizes 2B & 7B in resolutions 224 & 448 - ByteDance released larger SA2VA that comes in 26B parameters - Dataset: VRC-Bench is a new diverse benchmark for multimodal LLM reasoning performance
💬 LLMs - MiniMax-Text-01 is a new huge language model (456B passive 45.9B active params) by MiniMaxAI with context length of 4M tokens 🤯 - Dataset: Sky-T1-data-17k is a diverse dataset used to train Sky-T1-32B - kyutai released Helium-1-Preview-2B is a new small multilingual LM - Wayfarer-12B is a new LLM able to write D&D 🧙🏻♂️ - ReaderLM-v2 is a new HTML parsing model by Jina AI - Dria released, Dria-Agent-a-3B, new agentic coding model (Pythonic function calling) based on Qwen2.5 Coder - Unsloth released Phi-4, faster and memory efficient Llama 3.3
🖼️ Vision - MatchAnything is a new foundation model for matching - FitDit is a high-fidelity VTON model based on DiT architecture
🗣️ Audio - OuteTTS-0.3-1B is a new multilingual text-to-speech model with voice cloning and emotion control capabilities
📖 Retrieval - lightblue released a new reranker based on Qwen2.5 LB-reranker-0.5B-v1.0 that can handle 95+ languages - cde-small-v2 is a new sota small retrieval model by @jxm
🏎️ Today I'm introducing a method to train static embedding models that run 100x to 400x faster on CPU than common embedding models, while retaining 85%+ of the quality! Including 2 fully open models: training scripts, datasets, metrics.
We apply our recipe to train 2 Static Embedding models that we release today! We release: 2️⃣ an English Retrieval model and a general-purpose Multilingual similarity model (e.g. classification, clustering, etc.), both Apache 2.0 🧠 my modern training strategy: ideation -> dataset choice -> implementation -> evaluation 📜 my training scripts, using the Sentence Transformers library 📊 my Weights & Biases reports with losses & metrics 📕 my list of 30 training and 13 evaluation datasets
The 2 Static Embedding models have the following properties: 🏎️ Extremely fast, e.g. 107500 sentences per second on a consumer CPU, compared to 270 for 'all-mpnet-base-v2' and 56 for 'gte-large-en-v1.5' 0️⃣ Zero active parameters: No Transformer blocks, no attention, not even a matrix multiplication. Super speed! 📏 No maximum sequence length! Embed texts at any length (note: longer texts may embed worse) 📐 Linear instead of exponential complexity: 2x longer text takes 2x longer, instead of 2.5x or more. 🪆 Matryoshka support: allow you to truncate embeddings with minimal performance loss (e.g. 4x smaller with a 0.56% perf. decrease for English Similarity tasks)
Check out the full blogpost if you'd like to 1) use these lightning-fast models or 2) learn how to train them with consumer-level hardware: https://huggingface.co/blog/static-embeddings
The blogpost contains a lengthy list of possible advancements; I'm very confident that our 2 models are only the tip of the iceberg, and we may be able to get even better performance.
