Open-R1: a fully open reproduction of DeepSeek-R1

Hugging Face Blog · 1 year ago

A research team launched Open-R1 to reproduce DeepSeek-R1's training methods after DeepSeek released a reasoning model that matched OpenAI's o1 performance but kept the training code and datasets proprietary. The project will systematically reconstruct the data pipeline and reinforcement learning approach across three steps, starting with distilling datasets from DeepSeek-R1 and progressing to pure RL training on math, code, and reasoning tasks. This open-source reproduction aims to let any researcher fine-tune existing language models into reasoning models using the shared datasets and training recipes.