ByteDance and Tsinghua University introduced the DAPO training system

Teams from ByteDance Seed and Tsinghua University (AIR) have released an open-source reinforcement learning system called DAPO. The project focuses on optimizing the training policies of models and demonstrates high efficiency in mathematical reasoning tasks.

In testing, the system achieved a score of 50% on the AIME 2024 benchmark using the base model Qwen2.5-32B. The key technical features of the implementation are the decoupled clip (separated constraint) and dynamic sampling mechanisms, which allow for more stable model weight updates during training.

The project repository contains the source code necessary to reproduce the results and integrate the system into custom training pipelines. The developers emphasized reducing computational costs while maintaining high accuracy of logical inferences, which is critical for models working with complex mathematical tasks.

The implementation of DAPO enables researchers to apply more flexible optimization strategies than standard PPO methods. The use of dynamic sampling reduces the number of iterations required to achieve target accuracy metrics, making the system a promising tool for fine-tuning LLMs in specialized areas.