open-r1

Why we included this project

If you want to train your own reasoning model instead of just calling an API, this repo gives you a working place to start. It breaks the DeepSeek-R1 pipeline into concrete steps: supervised fine-tuning, GRPO reinforcement learning, synthetic data generation through Distilabel, and evaluation scripts, all tied together with a Makefile. The code is deliberately kept small and readable, so you can follow each script and swap in your own data or base model rather than fighting a heavyweight framework. The README also lays out the full plan for reproducing R1, from distilling a reasoning corpus down to pure RL training, which makes it a handy reference for how modern reasoning-model training fits together. Researchers and small teams get both runnable code and a clear roadmap, along with released datasets and training recipes such as OpenR1-Distill-7B.