You can deploy this model and start earning money today!

lewisdog/
qwen3-1.7b-cogs-ingest-GGUF

Name: qwen3-1.7b-cogs-ingest-GGUF
Availability: InStock

information-extraction

Views

No views yet

qwen3-1.7b-cogs-ingest — GGUF

GGUF build of lewisdog/qwen3-1.7b-cogs-ingest: the ingest student for the cogs knowledge engine — entity/claim extraction and summarisation of raw captures into a wikilinked markdown vault, small enough to run fully local (including in the browser via wllama).

Fine-tuned on 2,113 pairs distilled from 256 real vault ingests; replies in strict JSON under cogs-ingest's exact prompts (extract / suggest_links / page_update / contradiction).

Files

file	quant	size	notes
`qwen3-1.7b-cogs-ingest-Q8_0.gguf`	Q8_0	~1.8 GB	use this — see below
`qwen3-1.7b-cogs-ingest-adapter-f16.gguf`	LoRA adapter (f16)	~33 MB	apply onto a Qwen3-1.7B base with llama.cpp `--lora`

⚠️ Quantization sensitivity (measured)

This model's long structured extractions degrade sharply below 8-bit:

quant	behaviour
Q8_0	full extractions (summary + claims + entities) — matches the fp evaluation
Q6_K	summary only; claims/entities come back empty
Q4_K_M	schema-correct but all fields empty

Symptom of over-quantization is valid-but-empty JSON, not garbage — easy to miss in casual testing. Q4/Q6 builds are deliberately not published.

Serving notes (important)

Do NOT use a repetition penalty (--repeat-penalty 1.0). Extraction must copy tokens from the input document (titles, quotes, claim phrasing); a penalty suppresses exactly those tokens and empties the output. (This is the opposite of the sibling ask student, which needs 1.1.)
Disable Qwen3 thinking (chat_template_kwargs: {"enable_thinking": false}); trained non-thinking. Temperature 0.

Siblings

Ask student (grounded Q&A): lewisdog/qwen3-1.7b-cogs-ask-GGUF
This model merged (safetensors): lewisdog/qwen3-1.7b-cogs-ingest

lewisdog/qwen3-1.7b-cogs-ingest-GGUF