From GitHub · · 1m
GitHub - DAPO: Open-source RL System from ByteDance and Tsinghua
ByteDance and Tsinghua release open-source RL system DAPO achieving 50% on AIME 2024.
In brief
ByteDance and Tsinghua release open-source RL system DAPO achieving 50% on AIME 2024. ByteDance and Tsinghua open-source DAPO, an RL system that solves 50% of AIME 2024 problems using Qwen2.5-32B. Originally reported by GitHub.
DAPO Launch
ByteDance and Tsinghua release open-source RL system DAPO.
AIME 2024 Results
DAPO achieves 50 points on AIME 2024 using Qwen2.5-32B.
Key Improvements
DAPO uses token-level policy loss and dynamic sampling for better performance.
Training Details
DAPO trained on 17k math problems with 50% AIME 2024 accuracy.
Model Evaluation
DAPO-Qwen-32B solves complex math problems step by step.
Reproducibility
Full training code and datasets available for community use.
Next Steps
DAPO will be fully open-sourced on GitHub with detailed documentation.
Community Support
Researchers can submit issues for DAPO on GitHub for discussion.
At a glance
ByteDance and Tsinghua open-source DAPO, an RL system that solves 50% of AIME 2024 problems using Qwen2.5-32B.
