ByteDance Seed and Tsinghua AIR have released DAPO, a fully open-source system for large-scale reinforcement learning of large language models. The release includes the Decoupled Clip and Dynamic sAmpling Policy Optimization algorithm, code infrastructure, training and validation datasets, reproduction scripts, and model weights. The DAPO-Qwen-32B model is trained from Qwen2.5-32B and achieves 50 points on AIME 2024, according to the project, exceeding the cited DeepSeek-R1-Zero-Qwen-32B result while using half as many training steps. An earlier DAPO version without token-level policy-gradient loss and dynamic sampling achieved 44 points on the same benchmark. The project provides DAPO-Math-17k for training and AIME 2024 for validation, along with scripts for data preparation, training, inference, and evaluation using Ray Serve and vLLM. The repository also publishes training records and a checkpoint, and describes response length, reward, entropy, and mean-probability trends as signals used to monitor training stability. Reproduction experiments were conducted on the Volcano Engine Machine Learning Platform, with a fuller platform reproduction guide planned.
