56127 JANV01:53
articleHuggingFace Blogil y a 4 m.
Unlocking Agentic RL Training for GPT-OSS: A Practical Retrospective
Agentic reinforcement learning trains LLMs over multi-step interactions with an environment, optimizing on-policy data as the agent plans, uses tools, and refines its decisions. The post presents Verl as an OSS framework, experiments with GPT-OSS-20B (and 120B) and benchmarks like Qwen-2.5-32B, and uses ReTool to verify tool-assisted coding tasks. It also covers integrating Harmony's chat template and the end-to-end loop from rollout collection to policy updates (GRPO or PPO).