luminus-coder-30b-sft-r2-adapter

LoRA adapter (r=32, attention projections) for Qwen3-Coder-30B-A3B-Instruct, fine-tuned on verified code solutions including frontier problems. Load with PeftModel.from_pretrained(base_model, this_repo).

Eval (held-out set of 325 coding problems, k=4, unit-test verified)

	pass@1	pass@4
base (Qwen3-Coder-30B-A3B-Instruct), plain prompt	0.302	0.391
+ this adapter, with its intended prompting	0.359 (two seeds: 0.359, 0.359)	0.523 / 0.492

On a further held-out set of problems the base model had never solved, the adapter reaches pass@8 0.299 (k=8, T=0.7). Adapter provided as-is for reproducibility and comparison; the full experiment history lives in a private results repo.