Welcome to Open-R1 🐳🤗

Open-R1 is an open initiative to replicate and extend the techniques behind DeepSeek-R1, a state-of-the-art reasoning model, in a fully transparent and collaborative way: https://github.com/huggingface/open-r1

This organization is dedicated to:

• Sharing datasets and models built on the path to replicating DeepSeek-R1.
• Fostering meaningful discussions and collaboration in the Community tab.

By working together, we aim to create a robust foundation for reasoning models that the entire research and industry community can leverage.

Plan of attack

We are using the DeepSeek-R1 tech report as a guide to recreate their pipeline. The work can be broken down into three main steps:

• Replicate R1-Distill: Distill a high-quality reasoning corpus from DeepSeek-R1 to create the R1-Distill models.
• Recreate the pure RL pipeline: Reproduce the reinforcement learning process that DeepSeek used to train R1-Zero. This will likely require curating new, large-scale datasets for math, reasoning, and code.
• Demonstrate end-to-end training: Show that we can go from a base model to RL-tuned reasoning capabilities through a multi-stage training approach, combining supervised fine-tuning (SFT) and reinforcement learning (RL).

How to contribute

This project thrives on community participation! Here are some ways you can contribute:

• Join the discussion: Share ideas, ask questions, and collaborate with others in the Community tab.
• Contribute code or datasets: Submit pull requests with datasets, models, or improvements to the pipeline.
• Experiment and share results: Try out different approaches and share your findings with the community.

Let’s build something impactful together. 🚀