Robert Nishihara
@robertnishihara
Co-founder @anyscalecompute. Co-creator of @raydistributed. Previously AI research at Berkeley.
加入 March 2009
875 正在關注 17.3K 粉絲
An agentic RL training guide for Qwen 3.5 397B with SkyRL. They show very large improvements with roughly 2000 expert-labeled tasks.
How does one RL post-train a 397B model for long-horizon knowledge work? 👩💼
We share every step we took to bring Qwen 3.5 397B from 16.1% Pass@1 to 27.3% on APEX-Agents using DPPO, including final models weights and the full training script🚀 This is the first of many works from Mercor Research on open model training research.
Full blog:
Source code:
顯示更多
