Lewis Tunstall

Bestselling book co-authored with Leandro von Werra and Thomas Wolf, providing a hands-on guide to building NLP applications with Hugging Face Transformers. Revised edition published in full color.

Core contributor to Hugging Face's library for fine-tuning and aligning language models via RLHF, DPO, and other preference learning methods. One of the most widely used alignment libraries in the open-source ecosystem.

Led the Zephyr project demonstrating that distilled DPO (dDPO) with AI feedback can produce a 7B-parameter chat model surpassing Llama2-Chat-70B on MT-Bench, without requiring human annotation.

Open-source collection of robust recipes to align language models with human and AI preferences, covering SFT, DPO, Constitutional AI, and model-specific fine-tuning guides.

Efficient few-shot learning framework that fine-tunes Sentence Transformers contrastively on small labeled sets, achieving competitive accuracy with orders of magnitude fewer parameters than prompt-based methods.

Teamed up with Numina to create an 850k math problem-solution dataset and won the first AI Math Olympiad progress prize, demonstrating the power of high-quality open datasets for mathematical reasoning.

Co-authored SmolLM2 and SmolLM3 small language models and co-launched the Open-R1 initiative to reproduce open reasoning model training, including OlympicCoder for competitive programming.

Co-developed the free Hugging Face NLP course bridging software engineers into the Transformers ecosystem, covering tokenization, fine-tuning, and deployment across multiple modalities.