bernardw/qwen3-4b-codereview-taid-q4-mlc

Qwen3-4B CodeReview · TAID · q4f16_1 (WebLLM / browser build)

In-browser (WebGPU) q4f16_1 build of a Qwen3-4B student distilled for structured code review. Runs client-side via WebLLM / MLC-LLM — no server, no data leaves the machine.

What it is

• Student: Qwen3-4B, LoRA-distilled (forward-KL / TAID, native temperature, response-position).
• Teacher: Qwen3-8B (8-bit), targets generated with thinking disabled (clean JSON).
• Task: read a numbered code chunk, emit a JSON findings array (category, severity, line, evidence…).
• Quality: held-out planted-bug F1 ≈ 0.562 (bf16 eval) — the best student in this project, beating the earlier Qwen2.5-Coder-3B students (0.478 TAID / 0.509 best-single / 0.535 soup).
• Format: MLC q4f16_1, ~2.1 GB, 4.5 bits/param, context 4096, prefill chunk 1024.

Running in the browser

Pair these weights with WebLLM's prebuilt Qwen3-4B model library (web-llm-models/v0_2_84/base/Qwen3-4B-q4f16_1_cs1k-webgpu.wasm). Example model record:

{
  model: "https://huggingface.co/bernardw/qwen3-4b-codereview-taid-q4-mlc",
  model_id: "Qwen3-4B-CodeReview-TAID-q4f16_1",
  model_lib: "https://raw.githubusercontent.com/mlc-ai/binary-mlc-llm-libs/main/web-llm-models/v0_2_84/base/Qwen3-4B-q4f16_1_cs1k-webgpu.wasm",
  overrides: { context_window_size: 4096 }
}

License

Apache-2.0 (Qwen3 base is Apache-2.0; distilled from Qwen3-8B outputs). No Gemma terms apply.