DAPO
An open-source reinforcement learning system jointly developed by ByteDance's Seed team and Tsinghua University's AIR. It focuses on alignment and reasoning capability training for large language models, integrating components such as data processing, reward design, distributed training, and evaluation into a unified pipeline.
Timeline (last 90 days)
DAPO is an open-source reinforcement learning system jointly launched by ByteDance's Seed team and Tsinghua University's AIR (Institute for AI Industry Research), focusing on alignment and reasoning capability training for large language models.
Reinforcement learning based on verifiable rewards is the common starting point for recent works such as DeepSeek-R1 and DAPO.
The DAPO project attracted attention after being shared on Hacker News.
All Facts (2)
DAPO is an open-source reinforcement learning system jointly launched by ByteDance's Seed team and Tsinghua University's AIR (Institute for AI Industry Research), focusing on alignment and reasoning capability training for large language models.
50%UnverifiedReinforcement learning based on verifiable rewards is the common starting point for recent works such as DeepSeek-R1 and DAPO.
50%