Qwen3.5-122B-A10B-Banya-Tuned-v21-grpo

Option D3 + v10 (Masked SFT) init + dense reward — GRPO with multi-stage preflight reward.

• init: v10 LoRA (Masked SFT, assistant-only loss)
• 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, 80 steps, k=8
• train stats: REAL PASS 24/80 = 30%, train_loss 0.0855, train_runtime 18h 16m