Open-R1: Update #1

Hugging Face Blog · 1 year ago

The Open-R1 project, launched to replicate DeepSeek R1's training pipeline and synthetic data generation, has reproduced DeepSeek's evaluation scores on the MATH-500 benchmark within one week. The team scaled synthetic data generation from 2 H100 nodes to 32 GPUs across 4 nodes to handle DeepSeek R1's average response length of 6,000 tokens while maintaining stable GPU utilization through streaming inference rather than batched requests. The community has since built dozens of projects including distilled models, multimodal versions, and open datasets containing 17,000 to 800,000 reasoning examples for fine-tuning smaller models.