Open-R1: Open Reproduction of DeepSeek-R1

AI Impact Summary

The Open-R1 project aims to fully reproduce DeepSeek-R1, a reasoning model built on DeepSeek-V3, by reconstructing its training data and pipeline. This is critical for the open-source community to understand and replicate DeepSeek’s innovative reinforcement learning approach, particularly the use of Group Relative Policy Optimization (GRPO) and Multi Token Prediction (MTP). Replicating this model will allow for further experimentation and development of open reasoning models, addressing key questions around data curation, model training hyperparameters, and scaling laws.