CoreCoder_RL is a lightweight research project for agentic RL training built on slime and CoreCoder.
corecoder_rl/: CoreCoder RL proxy service, rollout, and automated training code.scripts/: training scripts.patches/corecoder_slime_overlay/: FSDP training overlay.
- The policy model samples on-policy responses through the CoreCoder RL proxy.
- A local or external model can optionally role-play the student and generate natural follow-up questions or feedback in teacher-student dialogues.
- The teacher observes the same generated student trajectory and computes
teacher_log_probs.
The CoreCoder_RL glue code in this repository is released under the MIT License. Upstream projects and copied/overlaid files remain governed by their original licenses. Please review the licenses of CoreCoder, slime, SGLang, Ray, Transformers, PyTorch, and the relevant models before redistribution or commercial use.
This project is built upon the excellent work of CoreCoder, OpenClaw-RL, slime, and ToolPRM. We thank the authors of these projects and papers.