rl_r2egym-nl2bash-swesmith-pymethods2test_terminus-structured

RL-trained Qwen3-8B with structured tool calls.

Training pipeline: SFT (r2egym+nl2bash+swesmith) → RL mixed dataset (37 steps) → RL full r2egym (55 steps) → RL pymethods2test (110 steps).

Key results:

• SWEBench-100: 42% pass@3 (vs 37% base with terminus-2)
• Pymethods2test: 94-100% pass@8
• 14 SWEBench tasks solved that base model cannot
• Trained with terminus-structured agent (bash, view, edit, create, search tools)

Training Details

• Base model: laion/r2egym-nl2bash-stack-bugsseq-fixthink-again
• Training method: rloo-n with terminus-structured agent (structured tool calls: bash, view, edit, create, search)
• Framework: BenSkyRL + Harbor
• Context: 32k (24k input + 8k output)
• Learning rate: 1e-5

SWEBench-Verified Results (100 tasks, pass@3)

Model	SWEBench pass@3
Base SFT (terminus-2)	37%
This model (terminus-structured)	See eval results

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("laion/rl_r2egym-nl2bash-swesmith-pymethods2test_terminus-structured")
tokenizer = AutoTokenizer.from_pretrained("laion/rl_r2egym-nl2bash-swesmith-pymethods2test_terminus-structured")