metadata
license: apache-2.0
base_model: laion/r2egym-nl2bash-stack-bugsseq-fixthink-again
tags:
- reinforcement-learning
- code
- pymethods2test
- terminus-structured
- rloo-n
RL Pymethods2test Fresh (Step 150) - Terminus Structured
RL-trained Qwen3-8B with terminus-structured agent on pymethods2test dataset.
- Base model: laion/r2egym-nl2bash-stack-bugsseq-fixthink-again (8B SFT)
- Training: 150 RL steps on pymethods2test (5000 tasks), fresh from SFT base
- Agent: terminus-structured (bash, view, edit, create, search, submit tools)
- Algorithm: RLOO-N, lr=1e-5
- Training pass@8: 100%
- SWEBench-100 n=1: 33%