Open R1 is a fully open-source reproduction of DeepSeek-R1, a reasoning-focused language model. The repository aims to democratize access to the R1 pipeline by providing open implementations of the training, inference, and evaluation components that were previously proprietary. The project is written in Python and serves as a collaborative effort to build and improve upon DeepSeek's reasoning model architecture.
The repository is organized around three main training and data generation scripts within the src/open_r1 directory. The grpo.py script implements Group Relative Policy Optimization training, sft.py handles supervised fine-tuning on datasets, and generate.py creates synthetic data using Distilabel. A Makefile provides convenient commands for executing each step in the R1 pipeline. The project follows a structured three-step plan based on DeepSeek-R1's technical report: first replicating the R1-Distill models through distillation, second reproducing the pure RL pipeline used for R1-Zero, and third demonstrating multi-stage training from base model to RL-tuned model.
Recent development milestones show significant progress. In May 2025, the project completed Step 1 by releasing the Mixture-of-Thoughts dataset containing 350k verified reasoning traces distilled from R1, spanning mathematics, coding, and science tasks. This enabled the release of OpenR1-Distill-7B, which replicates the reasoning capabilities of DeepSeek-R1-Distill-Qwen-7B. In March 2025, the CodeForces-CoTs dataset of 10k competitive programming problems with 100k solutions was released, along with the IOI24 benchmark featuring very difficult international olympiad problems. A 7B Qwen model trained on this dataset outperformed Claude 3.7 Sonnet on IOI24. Earlier releases included the OpenR1-Math-220k dataset in February 2025 with 220k traces on NuminaMath, and the initial training, inference, and evaluation pipeline implementations.
The repository supports distributed training via both DDP and DeepSpeed with ZeRO-2 and ZeRO-3 configurations, configured for 8 x H100 GPUs by default. For GRPO training, the project leverages TRL's vLLM backend for scaling across multiple nodes. The codebase includes support for code execution as a reward function during training, with integration for E2B and Morph sandbox providers to safely execute and verify generated code against test cases. This enables training on verifiable tasks like Codeforces problems where solutions can be automatically validated.
The repository shares contributors with vllm-project/vllm, sgl-project/sglang, and flutter/flutter, suggesting integration with inference optimization and broader development ecosystems. The project is classified across multiple domains including multilingual support, model fine-tuning, AI research, natural language understanding, and transfer learning, reflecting its comprehensive approach to open reasoning model development.
