Qwen3-32B-SweSmith-20step

RL-trained Qwen3-32B on SWEsmith.

Training Details

• Base model: Qwen/Qwen3-32B
• Training method: rloo_n
• Training data: 2,500 SWEsmith tasks
• Steps: 20 global steps
• Infrastructure: 16x4 GH200 GPU nodes (GCP), FSDP2 with TP=2 for inference engines (24 inference engines + 4 policy/ref nodes)
• Sandbox environment: Beta9/Beam containers for code execution

Training Curve

Metric	Step 1	Step 10	Step 20
Avg Raw Reward	0.031	0.041	0.072
Pass@8	0.141	0.156	0.234