Third-round LoRA adapter for Qwen3-Coder-30B-A3B-Instruct, reviewing Drupal PHP diffs and emitting structured JSON findings. v3 = v2's data + 14 synthetic access/logic-bypass contrastive pairs, aimed at v2's weakest spot.

Honest result: v3 ≈ v2. For deployment, use v2. On the held-out real-defect set the 14 synthetic pairs produced no measurable improvement (the v2→v3 delta is one extra catch and one extra false alarm — noise at n=32), and v3 gave back v2's perfect specificity. The lesson: small synthetic doses don't close the gap; real, objective positives do. v4 acts on that.

Results: base vs v2 vs v3 (real-defect eval, n=32)

Metric	Base	v2	v3
Verdict accuracy	~72%	78.1%	78.1%
Positive recall	87.5% (14/16)	56.2% (9/16)	62.5% (10/16)
Negative specificity	~56%	100%	93.8%
Category match	56.2%	43.8%	43.8%
Invalid JSON	0/32	0/32	0/32

Training data

v2's 498 rows + 14 access/logic-bypass contrastive pairs (gen_bypass_pairs.py, Drupal-expert-verified) = 526 rows. QLoRA r=16 on q/k/v/o, batch4+grad-ckpt, MAX_LEN=2048, 3 epochs, lr 2e-4. Full 3-way report with per-item detail ships in the project repo under docs/eval/dcr-qlora-v3-report.md.

Limitations

Same as v2, plus: v3 traded v2's 100% specificity for one false positive without a real recall gain, so it is not a recommended upgrade over v2. Real-defect recall remains ~60%; the access-bypass class is still largely missed. Keep a human in the loop; the model is one component of a hybrid pipeline (static analyzers + RAG + this adapter).

bartek-flp/qwen3coder-30b-dcr-lora-v3

Qwen3-Coder-30B-A3B — DCR (Drupal Code Review) QLoRA adapter, v3 (experimental)

Results: base vs v2 vs v3 (real-defect eval, n=32)

Training data

Limitations

bartek-flp/
qwen3coder-30b-dcr-lora-v3