DAPO: An Open-source RL System from ByteDance Seed and Tsinghua AIR
Summary
DAPO is an open-source reinforcement learning system for large language models developed by ByteDance Seed and Tsinghua AIR. The repository describes the Decoupled Clip and Dynamic Sampling Policy Optimization (DAPO) algorithm, training infrastructure, datasets (DAPO-Math-17k; AIME 2024 evaluation), model weights (DAPO-Qwen-32B), and reproducibility via open-source scripts and wandb records. The page includes setup instructions, inference examples, and details on experiments and environments, along with links to related resources and publications.