Open Reproduction of DeepSeek-R1
Summary
Open Reproduction of DeepSeek-R1 is a fully open reproduction project hosted on HuggingFace/GitHub. The README outlines a plan of attack, installation steps, training models (SFT and GRPO), data generation, data decontamination, evaluation pipelines, and how to reproduce DeepSeek-R1 results. It highlights multi-node GPU training, vLLM backends, dataset handling, and code structure for training and inference, inviting community contributions.