rl_pymethods2test-r2egym_terminus-structured

RL-trained Qwen3-8B with structured tool calls (terminus-structured agent).

Training Pipeline

SFT (r2egym+nl2bash+swesmith) → RL mixed dataset (37 steps) → RL full r2egym (55 steps) → RL pymethods2test (156 steps, full epoch)

Key Results

• SWEBench-100 pass@3: 37-42% (depending on eval run)
• Pymethods2test pass@8: 91-97%
• SWEBench in-train eval: up to 14 fully solved at various checkpoints
• Training on test-writing (pymethods) maintains code-editing ability (SWEBench)

Training Details

• Base model: laion/r2egym-nl2bash-stack-bugsseq-fixthink-again
• Agent: terminus-structured (bash, view, edit, create, search tools)
• Algorithm: RLOO-N
• Learning rate: 1e-5
• Context: 32k (24k input + 8k output)
• Framework: BenSkyRL + Harbor (JSC HPC)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("laion/rl_pymethods2test-r2egym_terminus-structured")
tokenizer = AutoTokenizer.from_pretrained("laion/rl_pymethods2test-r2egym_terminus-structured")