DAPO

DAPO: An Open-Source LLM Reinforcement Learning System at Scale

Qiying Yu et al. · arXiv 2025

Read paper · arXiv:2503.14476v1

Improve policy optimization with informative-group filtering, asymmetric clipping, and token-level aggregation.

What you will implement

Implement group filtering and the token-level objective; sampling infrastructure and distributed training are outside this module.

• 13. DAPO Dynamic Filtering and Token Objective

  Retain informative groups and implement asymmetric token-normalized clipping.

  Two bounded core functions. Oversampling orchestration, overlong reward shaping, and distributed training are not reproduced.