hintgemma-dpo-q4f16_1-MLC

hintgemma-dpo (gemma-2-2b-it fine-tune) — q4f16_1 MLC build for WebLLM

A LoRA fine-tune of gemma-2-2b-it that writes Socratic mathematics hints for STACK questions without revealing answers, packaged for in-browser use with WebLLM. Built for the Moodle plugin local_stackhinter (the plugin's "custom on-device model" setting), where the model runs entirely in the student's browser: no API key, and the teacher answer never leaves the server.

Training & evaluation (2026-07-02)

• SFT on 1,994 hints from a 9,252-hint evaluation corpus (3-judge panel mean ≥ 4/5; zero flags from a substring leak detector, judge reveal flags, and a hand-adjudicated verbalised-leak audit), then anti-leak DPO on 93 preference pairs whose rejected side is a real observed answer leak.
• On the identical 1,236-hint harness that measured the base model: primary-judge quality 2.674 → 3.326 (paired Δ +0.638, 95% CI [+0.567, +0.715]) with 0/1236 literal leaks and 0 adjudicated verbalised leaks — the safe tier at above-every-stock-challenger quality.
• Details: the stack-question-forge research paper (§5.6) and research/data/verbal_leak_adjudication.json in that repository.

Using with WebLLM

This is a q4f16_1 conversion (context 4096, prefill 2048, gemma_instruction template) that reuses the prebuilt gemma-2-2b-it model library:

const appConfig = {model_list: [{
  model: "https://huggingface.co/dcregg/hintgemma-dpo-q4f16_1-MLC",
  model_id: "hintgemma-dpo-q4f16_1-MLC",
  model_lib: "https://raw.githubusercontent.com/mlc-ai/binary-mlc-llm-libs/main/web-llm-models/v0_2_84/base/gemma-2-2b-it-q4f16_1_cs1k-webgpu.wasm",
  overrides: {context_window_size: 4096},
}]};
const engine = await webllm.CreateMLCEngine("hintgemma-dpo-q4f16_1-MLC", {appConfig});

License / provenance

Derivative of Gemma (Google), distributed under and subject to the Gemma Terms of Use. Fine-tuning data was synthetic mathematics-tutoring text generated and judged by open models; no student or personal data.