ByteDance Seed and Tsinghua AIR released DAPO, an open-source large-scale reinforcement learning system featuring decoupled Clip and Dynamic Sampling Policy Optimization. Using a Qwen2.5-32B base model, the system achieved 50% on the AIME 2024 benchmark. The release includes algorithms, code infrastructure, and datasets built on the verl framework.
HOW THIS AFFECTS YOU
●
builderYou can use this open-source infrastructure to scale reinforcement learning for reasoning tasks.
●
researcherThe decoupling of Clip and Dynamic Sampling offers a new method for optimizing large-scale LLM RL.