Hugging Face Research

The science team at Hugging Face is dedicated to advancing machine learning research in ways that maximize value for the whole community. Our work focuses on three core areas of tooling, datasets and open models:

🛠️ Tooling & Infrastructure

The foundation of ML research is tooling and infrastructure and we are working on a range of tools such as datatrove, nanotron, TRL, LeRobot, and lighteval.

📑 Datasets

High quality datasets are the powerhouse of LLMs and require special care and skills to build. We focus on building high-quality datasets such as no-robots, FineWeb, The Stack, and FineVideo.

🤖 Open Models

The datatsets and training recipes of most state-of-the-art models are not released. We build cutting-edge models and release the full training pipeline as well fostering more innovation and reproducibility, such as Zephyr, StarCoder2, or SmolLM2.

🌸 Collaborations

Research and collaboration go hand in hand. That's why we like to organize and participate in large open collaborations such as BigScience and BigCode.

⚙️ Infrastructre

The research team is organized in small teams with typically <4 people and the science cluster consists of 96 x 8xH100 nodes as well as an auto-scalable CPU cluster for dataset processing. In this setup, even a small research team can build and push out impactful artifacts.

📖 Educational material

Besides writing tech reports of research projects we also like to write more educational content to help newcomers get started to the field or practitioners. We built for example the alignment handbook, the pretraining tutorial, or the FineWeb blog.