wp-qwen3-30b-a3b-wp-gen-v1.2

wp-qwen3-30b-a3b-wp-gen-v1.2 — WordPress code generation (Qwen3-30B-A3B)

Generation half of a two-model WordPress pair fine-tuned from Qwen3-30B-A3B. This model writes WPCS-compliant PHP from a plain-English instruction. The paired review/judge model lives at iamchum/wp-qwen3-30b-a3b-wp-judge-v1.3-gguf — route generation requests here with <wp_gen> and review requests there with <wp_judge>.

This card documents the full v3.0 compression lineage, including the stages that returned nothing. Two of them did, and that's recorded honestly rather than hidden.

The pair

• Base: Qwen3-30B-A3B (MoE, 128 experts, top-8, ~30.5B total / ~3.3B active, 48 layers).
• This repo ships the bf16 checkpoint: 57 GB, 13 safetensors shards, plus tokenizer (tokenizer.json carries the <wp_gen> / <wp_judge> task-token vocabulary).
• No quantized tier is shipped for the generation model — the Q8 evidence below applies to the judge model specifically (see the paired GGUF repo). bf16 is the only ship tier here.

Compression lineage (base -> RL -> MoE-Sieve -> merge -> prune -> quantize)

Each arrow is a real gate, applied to the pair as a whole. Scores are vLLM-served unless noted.

1. Base — Qwen/Qwen3-30B-A3B.
2. SFT (v1.2 / v1.3) — reasoning SFT for generation (v1.2, this repo), relabel-SFT for the judge (v1.3, rho 0.827 Tinker-runtime / 0.8017 vLLM single-seed).
3. RL (GSPO) — REJECTED. Warm-started from v1.3 with an oracle-passed calibration reward; killed on 6/6 dead checkpoint reads (2026-07-05). No RL checkpoint was promoted.
4. MoE-Sieve (expert drop) — no compression. Routing profile shows ~88-99 effective experts/layer of 128; every masked-k budget cuts live capacity. wp-bench collapses 0.4484 (full) -> 0.2275 (k=64) -> 0.0546 (k=32). optimal_k = full; nothing dropped.
5. LoRA merge — adapters merged into base weights; merged output matches adapter-on-base. This repo ships the merged checkpoint.
6. AIMER / REAP pruning — no winner. Weight-norm AIMER at the lightest ratio (25%) collapses generation to wp-bench 0.1577 and judge ensemble rho to 0.1651 (parse 44.6%). 50/75% skipped, REAP conditional-skipped per the pre-registered rule. Ships unpruned at full 128-expert width.
7. Quantization — measured on the judge model (see the paired GGUF repo); uniform 4-bit nf4 is a failure, Q8 is the recommended ship tier for the judge. This generation model ships bf16.

Net: the v3.0 pipeline confirmed the model cannot be shrunk by expert-count or weight-norm methods on this workload. Quantization is the only size reduction available, and it was applied to the judge (this generation repo ships full-precision bf16).

Usage (vLLM)

from vllm import LLM, SamplingParams

llm = LLM(model="iamchum/wp-qwen3-30b-a3b-wp-gen-v1.2", trust_remote_code=True, dtype="bfloat16")
sp = SamplingParams(temperature=0.0, max_tokens=1024)

prompt = "<wp_gen> Write a WordPress function to safely query posts by meta value."
out = llm.generate([prompt], sp, chat_template_kwargs={"enable_thinking": False})
print(out[0].outputs[0].text)  # WPCS-compliant PHP

Or via vllm serve iamchum/wp-qwen3-30b-a3b-wp-gen-v1.2 --trust-remote-code and the OpenAI-compatible /v1/chat/completions endpoint, sending chat_template_kwargs: {"enable_thinking": false} per request. <wp_gen> is an ordinary prompt-prefix string, not a special tokenizer control token — just prepend it to the instruction; the shipped tokenizer.json / chat_template.jinja handle the rest.

For review/critique of existing code, send the same prompt shape with <wp_judge> to the paired judge repo instead.

Benchmarks

Fresh, receipt-backed measurements on the shipping stack (vLLM bf16, this checkpoint, temperature 0.0), taken 2026-07-11.

wp-bench. The fresh full-suite score (0.4365) sits 1.19pp below the 0.4484 Gate-1 reference figure, well inside the project's measured 5.20pp seed-noise floor, and clears the 0.4286 acceptance bar. Same stack both sides (vLLM bf16, identical checkpoint and sampling); no regression signal.

SWE-bench, and why the number is low. SWE-bench Lite is Python-repository patch generation; this model is WordPress/PHP-specialized. A low out-of-domain number is expected and is not a quality defect: the model was never trained to patch Django or sympy, and the score is published for honest positioning, not vanity. Protocol: one oracle-retrieval prompt in, one unified-diff patch out (no agent scaffold, no retries), temperature 0.0, seed 0, max_model_len 24576, official swebench 4.1.0 containerized harness built natively for arm64. The scope (Lite-300 + PHP-43) was pre-registered before any result existed.

Full-scope denominators count everything against the model: of Lite's 300, 80 prompts exceeded the 24k serving context (scored unresolved), 1 generation was unparseable, 59 patches failed to apply, and 29 instances could not be evaluated because their 2018-era Python environments cannot build on arm64/2026 toolchains (also scored unresolved, conservative against the model). On the 131 instances that ran to a test verdict the resolved rate is 3.82% (5/131). The PHP-Multilingual subset is in-language but still out-of-domain (framework libraries, not WordPress): 0/43 full-scope, 0/20 evaluated; 17 over-length, 6 apply-failures.

Evaluation

• wp-bench (WordPress code correctness/knowledge/quality): generation 0.4484 (bar 0.4286); fresh full-suite rerun on the shipping stack 0.4365 (within seed noise — see Benchmarks).
• The pair's judge rho (Spearman vs human-relabeled 9-dim scores) is measured on the paired judge repo: ensemble 0.8075, single-seed 0.8017.

Limitations

• Two-model pair, not a single unified model. Both share the base; route by task token.
• No RL enhancement (rejected). No pruning speedup (no viable prune). This generation model ships bf16 only — no quantized tier was evaluated for generation.
• A materially better pair needs a stronger base (Qwen3.6-class), which is the intended next iteration of this exact pipeline.

Provenance

Full training/eval history and both negative pruning results are documented in the project repository: JOURNAL.md, PIPELINE.md, output/prune/prune_methodology.md, output/eval3/eval3_final_comparison.json.