Qwen3.5-122B-A10B-Banya-Tuned-v20-grpo

Qwen3.5-122B-A10B-Banya-Tuned-v20-grpo

Option D3 + dense reward + v5 init — GRPO with multi-stage preflight reward.

• init: v5 LoRA (mix corpus, ~30% Pass@1 baseline)
• trainer: TRL GRPOTrainer
• rollout: HF model.generate (k=8 per task, T=1.0)
• reward: dense [0,1.0] = parse 0.05 + grep 0.05 + file 0.10 + func 0.10 + harness 0.30/0.70
• MoE safeguards: output_router_logits + aux loss + explicit router freeze (from v19)
• corpus: SWE-bench-Lite 270 train pool (no leakage with stratified-30 eval)
• hyperparams: β=0.1, ε=0.2, lr=1e-6, 100 steps, k=8

Builds on v19 (GRPO + MoE safeguards validated stable for 21.5h, 8/30 smoke). v20 addresses v19's plateau by densifying reward signal (parse/grep/file/func preflight gives gradient even when harness is stuck at 0.3 ceiling).