lfm2.5-350m-cogs-ask-GGUF

lfm2.5-350m-cogs-ask

A 350M ask student for Cogitarium retrieval-QA. Two tasks:

• decompose: split a question into 1-4 retrieval sub-questions → {"subquestions":[...]}
• synth: answer strictly from provided wiki notes with inline [note-id] citations → {"answer","citations","abstained"}

Trained on the real cogs serialization — slash note-ids (concepts/planner) and [[wikilinks]] in note bodies — plus frontier-teacher-distilled grounded Q&A generated over the real vault. Correct on the deployment distribution; needs no serving hacks.

Serving pins

• temperature 0, repeat_penalty 1.0.
• Emits exact slash note-ids and handles [[wikilinks]] in bodies natively — you do NOT need to strip wikilinks or fuzzy-match citations (both were required by the earlier hyphen-id variant). Feed evidence as `### [note-id] Title

<body>` exactly as `cogs ask` builds it. - **Gate abstention upstream.** The `abstained` flag is only moderately reliable; decide "is this answerable?" from retrieval score / the decompose step and don't depend on the field.

Eval (real vault serialization; grounded citation validity)

Distilled from ~600 frontier-teacher-generated grounded Q&A over the real vault (plus the re-serialized base set). Grounded-citation validity:

Progression as teacher data scaled (out-of-domain grounded strict): vault-aligned only 38% → +204 teacher 50% → +570 teacher 58%. On the vault it is actually trained over, grounded citation is ~89%. Data quality/quantity — not model size — was the binding constraint. Feed evidence as `### [note-id] Title

<body-with-wikilinks>; citations come back as exact slash-ids, no serving hacks. Gate abstention upstream (the abstained` flag is only moderately reliable). A strong fast tier; the Qwen3-1.7B student remains the quality tier.

Recommended quant: Q8_0 (379 MB, 509 tok/s on GB10). decompose is flawless at any quant incl. Q4_K_M (229 MB).

This repo

GGUF quant ladder for llama.cpp (F16 / Q8_0 / Q4_K_M). See the quant table above for the recommended pick and per-quant behavior. Run with --temp 0 and the serving pins noted above.

Base model & license

Fine-tuned from LiquidAI/LFM2.5-350M. Use is governed by the LFM Open License v1.0 (lfm1.0) — see the LICENSE in the base repo. This derivative complies with and inherits those terms; attribution to LiquidAI is retained above.

Provenance

LoRA SFT (TRL) on the Cogitarium distillation datasets, DGX Spark (GB10). Full methodology, loss curves, eval harnesses and per-quant results: see the project RESULTS.md. This is the "fast/small tier" of the Cogitarium model picker; the Qwen3-1.7B students remain the quality tier.