sprapp/firstaid-openbook-360m-ternary

Sprapp open-book first-aid reader (ternary, on-device)

⚠️ Not medical advice. This is a research/demo model for grounded first-aid information. It is not a substitute for professional help. In an emergency, call your local emergency number immediately.

A ternary (≈1.58-bit) open-book reader, fine-tuned from SmolLM2-360M to answer only from a provided context (a retrieved first-aid passage) and to abstain when the answer isn't there. Runs on-device / offline in the Sprapp WASM engine — the use case is offline emergency help with no internet.

Pipeline: question → retrieve relevant passage from a first-aid knowledge pack (on-device BM25/FTS) → feed as <context> to this reader → grounded answer + safety disclaimer.

Attribution (Apache-2.0)

Derivative of SmolLM2-360M (HuggingFaceTB), Apache-2.0. License retained and base model credited.

Honest limitations

• Grounding is imperfect. At 360M and a small knowledge pack, the model tends to memorize topics rather than strictly read the provided context, so it can blend or drift between first-aid topics. It always emits the emergency disclaimer and the retrieved source is shown, but do not treat its wording as authoritative.
• Trained on a curated 12-topic first-aid pack (bleeding, burns, choking, CPR, sprains, etc.) — narrow scope.

Files

file	what
model_openbook.knm	ternary weights, KNM1 v3 (~220 MB)
tokenizer.json	SmolLM2 tokenizer
firstaid_kb.jsonl	the 12-topic first-aid knowledge pack (for the retriever)

Part of the Sprapp project. Built on SmolLM2 (Apache-2.0).