Mustafa Eyceoz

Simplify LLM post-training with the Training Hub library, which provides a common, pythonic interface for running language model post-training algorithms.

Dive into LLM post-training methods, from supervised fine-tuning and continual learning to parameter-efficient and reinforcement learning approaches.

Discover Async-GRPO, a new library for reinforcement learning tasks that efficiently handles large models, eliminates bottlenecks, and accelerates experiments.

Discover how the adaptive SVD approach enables LLMs to continually learn and adapt without forgetting previously acquired knowledge.

Learn about an efficient inference scaling method that can improve your model's reasoning ability and performance at runtime while saving on compute costs.

Progress in small LLM reasoning: Our Qwen-32B model, using particle filtering, now surpasses o1-preview on Math500.

On reproducing R1-like reasoning in small LLMs: LIMO dataset ineffective for Llama/Granite; synthetic data generation shows promise but fine-tuning is tricky.

An update on reproducing R1-like reasoning in small LLMs: Granite models show big gains with particle filtering, outperforming GPT-4o on benchmarks.