---

license: apache-2.0

base_model: Qwen/Qwen3.6-35B-A3B

pipeline_tag: text-generation

library_name: gguf

language:

- en

tags:

- wordpress

- php

- code-review

- moe

- qwen3

- gguf

---

WP Judge v4 (Qwen3.6-35B-A3B)

An open-weight WordPress code reviewer. Give it a PHP function, it returns a structured critique

scored across 9 WordPress-quality dimensions (WPCS compliance, SQL safety, security, performance, WP API

usage, code quality, dependencies, i18n, accessibility) and tells you what is wrong and why.

It judges, it does not generate. For WordPress code generation, use the current base model directly —

Qwen/Qwen3.6-35B-A3B with prompt-side task framing. No

generation fine-tune is shipped alongside this judge.

Acquisition

Repo: iamchum/wp-qwen3.6-35b-a3b-wp-judge-v4-gguf

| File | Quant | Size | MTP / speculative decoding |

|---|---|---|---|

| wp-judge-v4-unpruned.Q5_K_M.gguf (recommended) | Q5_K_M | 23.61 GiB | yes |

| wp-judge-v4-pruned-k224.Q6_K.gguf | Q6_K | 23.47 GiB | no |

Both judge identically within measurement noise (see Performance). Take the recommended file —

it supports speculative decoding (faster serving) at the same download size to within 150 MB.

The pruned file is for operators who want the marginally smaller expert footprint (224/256

experts) and don't need speculation.

hf download iamchum/wp-qwen3.6-35b-a3b-wp-judge-v4-gguf \
  wp-judge-v4-unpruned.Q5_K_M.gguf --local-dir ./judge

Use

Serve with llama.cpp (b9180+):

llama-server -m ./judge/wp-judge-v4-unpruned.Q5_K_M.gguf --host 127.0.0.1 --port 8020 \
  -ngl 999 -c 16384 --jinja -a wp-judge-v4 --parallel 4 --spec-type draft-mtp

--spec-type draft-mtp enables speculative decoding via the model's own MTP head (56% measured

draft acceptance — no separate draft model). Omit it (or use the pruned file) if you don't want it.

--jinja is load-bearing — the chat template drives the rubric output format.

The pruned file ignores --spec-type draft-mtp (no MTP head) but is otherwise served identically.

Ollama (pulls the GGUF straight from this repo):

ollama run hf.co/iamchum/wp-qwen3.6-35b-a3b-wp-judge-v4-gguf:Q5_K_M

Three things matter on Ollama: set the context window — the default truncates long critiques mid-rubric

(/set parameter num_ctx 16384 in the session, or PARAMETER num_ctx 16384 in a Modelfile) — note

that Ollama does not expose llama.cpp's --spec-type flag, so MTP speculative decoding is unavailable

there (both files serve at the same speed) — and disable thinking mode. Ollama's Qwen3.6 chat template

opens a <think> block that the judge's output format never closes, so with thinking enabled (the

default) the entire judgment — reasoning and <judge_output> verdict — arrives in message.thinking

and message.content comes back empty. Pass "think": false in /api/chat requests (or read the

thinking field instead); verified on Ollama that the full output then lands in content. The chat

template is read from the GGUF metadata.

All published quality numbers were measured on llama.cpp; other engines (Ollama, LM Studio,

llama-cpp-python) run the same GGUF but their numerics are unverified here.

Judge a snippet — prepend the <wp_judge> task prefix and the model returns per-dimension reasoning

followed by a structured verdict:

curl -s http://127.0.0.1:8020/v1/chat/completions -d '{
  "model": "wp-judge-v4",
  "messages": [{"role": "user",
    "content": "<wp_judge> Evaluate this WordPress code:\n\n```php\n<?php $wpdb->query(\"SELECT * FROM wp_posts WHERE ID=$id\"); ?>\n```"}],
  "temperature": 0.0, "max_tokens": 2048
}'
# -> reasoning per dimension, then a <judge_output> JSON block:
# {"verdict": "FAIL", "wpcs_compliance": 9, "security": 2, "sql_safety": 1, ..., "overall_score": 41}

<wp_judge> is a plain-text prompt prefix, not a special token. Dimensions that don't apply to a given

snippet (e.g. accessibility on a REST handler) are omitted from the JSON rather than scored.

Performance

Spearman rho vs. human-relabeled 9-dimension scores, n=121 held-out items, single-seed:

| Stack / config | rho | Notes |

|---|---|---|

| f16 GGUF, llama.cpp | 0.8002 | uncompressed floor |

| Q6_K GGUF, llama.cpp (shipped) | 0.8063 | 0 parse failures |

| Q8_0 GGUF, llama.cpp | 0.7851 | 0 parse failures |

All three rungs are statistically indistinguishable at this sample size (95% CI half-widths ~7-8pp). Q6_K

ships as the smallest tier with zero parse failures, not because it scored highest.

The unpruned variant measured rho 0.8093 (Q5_K_M, same protocol, 0 parse failures) against its own f16

floor of 0.8081 — statistically indistinguishable from the pruned file's numbers above. Its quant tier was

selected by the same rule: smallest rung within ±2pp of its own f16 floor with zero parse failures.

The prior WP Judge (iamchum/wp-qwen3-30b-a3b-wp-judge-v1.3-gguf, Qwen3-30B-A3B base, 30.2 GiB) reports

0.8056 on a 3-seed ensemble — a different stack and seed configuration than the single-seed number

above, and not directly comparable to it.

Variants

| | wp-judge-v4-unpruned.Q5_K_M.gguf (recommended) | wp-judge-v4-pruned-k224.Q6_K.gguf |

|---|---|---|

| Experts | 256/256 (unpruned) | 224/256 (expert-pruned) |

| MTP / speculative decoding | available (--spec-type draft-mtp, 56% measured draft acceptance) | not available |

| Size | 23.61 GiB | 23.47 GiB |

| Judge quality | statistically tied (see Performance) | statistically tied (see Performance) |

Why the pruned file has no MTP head: expert-pruning left the MTP layer at a different expert count than

the trunk, and GGUF's expert-count metadata must be uniform, so the MTP tensors were dropped at its

conversion (--no-mtp). Standard generation and judging are unaffected in both files. The two files are

statistically indistinguishable on judge quality; the recommendation is purely serving speed.

Links

• Base: fine-tuned from Qwen/Qwen3.6-35B-A3B, then quantized to GGUF. The recommended file keeps all 256 experts (MTP intact); the pruned variant is additionally expert-pruned to 224/256.
• How this was built: github.com/dr-robert-li/wp-finetune — see PIPELINE.md, JOURNAL.md, CHANGELOG.md.
• License: Apache 2.0.