DCR v6 — GGUF (Q4_K_M) for Ollama
A merged, quantized build of the DCR v6 Drupal code reviewer, ready for Ollama / llama.cpp on a Mac or a single 24GB GPU. It is the v6 fine-tune of Qwen/Qwen3-Coder-30B-A3B-Instruct (warm-started v4 → v5.1 → v6), merged into the base and quantized to Q4_K_M (~18.6 GB).
For the vLLM/LoRA form (NVIDIA servers), use the adapter:
bartek-flp/qwen3coder-30b-dcr-lora-v6.
Run it (Ollama)
1ollama pull hf.co/bartek-flp/qwen3coder-30b-dcr-v6-GGUF:Q4_K_M
2ollama cp hf.co/bartek-flp/qwen3coder-30b-dcr-v6-GGUF:Q4_K_M dcr-v6
Then point the Drupal CodeReviewer at --model dcr-v6.
How v6 did
100-pair clean eval (buggy pre-fix diff + matched merged fix, none seen in training), scored pair-wise — a pair counts only if the model flags the buggy half and stays quiet on the clean one:
| model | pair-correct | positive recall | negative specificity | category |
|---|
| base | 0.30 | 0.38 | 0.90 | — |
| v4 | 0.53 | 0.55 | 0.96 | 0.69 |
| v5.1 | 0.67 | 0.72 | 0.93 | 0.65 |
| v6 | 0.73 | 0.78 | 0.94 | 0.80 |
v6 is best on every metric, with a clean monotonic trend across versions. Honest note: the v6-over-v5.1 gap is directional but not statistically significant on 100 pairs (sign-test p ≈ 0.21) — a real, consistent improvement, not a slam-dunk. Full write-up lives in the Drupal CodeReviewer repo (docs/eval/v6-results.md).
What it catches
Drupal idioms that break things (missing DI, absent cache metadata, procedural hook that should be #[Hook]), logic defects (null deref, dropped guard, off-by-one, foreach-by-ref), and security (XSS, SQLi, access bypass, unsafe deserialization). Output is clean JSON findings with severity, category, file:line, and a fix.
License
Weights for research and evaluation. Commercial use of the surrounding Drupal CodeReviewer system needs a license:
filipiuk.bartek@gmail.com.
