Open Reproduction of DeepSeek-R1

Hugging Face releases open-source reproduction of DeepSeek-R1 reasoning model, completing step one of three-stage plan.

The open-r1 project provides scripts and datasets to reproduce DeepSeek-R1's reasoning capabilities. Step 1 completion (May 2025) released Mixture-of-Thoughts, a 350k-sample reasoning dataset distilled from R1 covering mathematics, coding, and science tasks. The project includes training recipes for OpenR1-Distill-7B using supervised fine-tuning and group relative policy optimization (GRPO), plus data generation pipelines. Results show OpenR1-Distill-7B matching or exceeding DeepSeek-R1-Distill-Qwen-7B on several benchmarks. Steps 2 and 3 involve reproducing the pure RL pipeline and demonstrating multi-stage training from base model. The codebase supports distributed training via DeepSpeed, vLLM inference, and code execution environments for competitive programming tasks.