qwen3-1.7b-cogs-ask-GGUF

qwen3-1.7b-cogs-ask — GGUF

GGUF build of lewisdog/qwen3-1.7b-cogs-ask-lora merged onto Qwen/Qwen3-1.7B: the ask student for the cogs knowledge engine — grounded, cited question answering over a personal markdown vault, small enough to run fully local (including in the browser via wllama).

What it does

Fine-tuned on 538 pairs distilled from cogs' real ask pipeline (question decomposition + grounded synthesis over retrieved notes), including deliberately unanswerable probes so the model learns to abstain rather than draw on outside knowledge. It expects cogs-ask's exact prompts and replies in strict JSON:

• decompose → {"subquestions": ["..."]}
• synthesize → {"answer": "... [note-id] ...", "citations": ["note-id"], "abstained": false}

Files

Serving notes (important)

• Decode with a small repetition penalty (--repeat-penalty 1.1 / penalty_repeat: 1.1): pure greedy can run away on long list outputs.
• Disable Qwen3 thinking (chat_template_kwargs: {"enable_thinking": false} or prefix an empty think block): the student was trained non-thinking.
• Temperature 0 recommended for schema-exact JSON.

Validated post-quantization (browser, wllama v3.5.1): decompose ✓, grounded synthesis with citation validation ✓ (~0.9s), abstention on out-of-domain ✓.

Siblings

• LoRA adapter: lewisdog/qwen3-1.7b-cogs-ask-lora
• Ingest student (extraction/summarisation): lewisdog/qwen3-1.7b-cogs-ingest