ByteDance Seed and Tsinghua AIR have released DAPO, an open-source reinforcement learning system for large-scale LLMs. The system achieves 50+ points on AIME 2024 using Qwen2.5-32B, outperforming prior methods with fewer training steps. DAPO introduces Decoupled Clip and Dynamic Sampling Policy Optimization. It includes algorithm, code, dataset, and model weights.

Training records show improved length stability, reward score consistency, and entropy control. Environment setup uses conda and pip. Inference code and example problems are provided.

The project builds on the verl framework. Researchers are invited to discuss via GitHub issues.