VictorChenCA/lokol-health-qwen3-1.7b-gguf

Lokol Health 1.7B (GGUF, Q4_K_M)

A LoRA fine-tune of Qwen/Qwen3-1.7B for Lokol Health: an offline assistant for nurse aides in rural Solomon Islands, in Solomon Islands Pijin and English, grounded in the Ministry of Health's Standard Treatment Manual for Children (2017). Built with Lokol Studio for Hack-Nation 7 x World Bank "Small AI for Development" (Health track).

Not a medical device. Not for clinical use. A prototype trained on synthetic data and not validated by clinicians.

What it does

Given a nurse's message, flags ([lang=pis] [rdt=yes] [act=yes] [transport=next_boat]) and one retrieved manual excerpt ([guideline: SECTION pN] ...), it answers in a fixed protocol: ``` ACTION: ADVISE | REFER_NOW | REFER_NEXT_TRANSPORT | ASK_PERSON STM: | NONE

<at most 6 short lines in the nurse's language>

It never diagnoses. In the Lokol app and bridge, every reply passes a rule-based safety gate (danger signs force referral, no guideline
means "ask a person", and no dose is shown unless the cited page gives it for that drug).

## Run it in one line (Ollama)
```sh
ollama run hf.co/VictorChenCA/lokol-health-qwen3-1.7b-gguf

Ollama reads template, system and params from this repo: Qwen3 ChatML with thinking off (as in training), the Lokol system prompt, temperature 0.2 and a 4k context. Every message is three lines in the Lokol protocol: flags, one manual excerpt, the nurse's message. Wrap multi-line input in triple quotes: text >>> """[lang=pis] [rdt=yes] [act=yes] [transport=next_boat] ... [guideline: FEVER p40] Fever is defined as an axillary temperature greater than 37.5 C. INVESTIGATIONS: consider blood film or rapid diagnostic test for malaria parasites. MANAGEMENT: treat the cause of the fever; consider giving paracetamol if fever above 38 C AND the child is irritable: dose 15 mg/kg (max 1 g) every 4 to 6 hours (maximum 4 doses in 24 hours); do not use aspirin in children for fever; give extra fluids. ... Pikinini 3 yia, 13 kilo, hot bodi fo tufala dei, hem kros an no laek kaikai tumas. Wanem mi mas duim?""" Expected shape of the reply: ```text ACTION: ADVISE STM: FEVER

<up to 6 short lines in Pijin: danger signs to check, RDT, paracetamol by weight, no aspirin, fluids, when to come back>

Use `[guideline: none]` when the manual has nothing on the question; the model should answer `ACTION: ASK_PERSON`.
To serve the Lokol bridge (WhatsApp, Messenger) from Ollama:
`LLM_URL=http://127.0.0.1:11434/v1/chat/completions LLM_MODEL=hf.co/VictorChenCA/lokol-health-qwen3-1.7b-gguf .venv/bin/python -m bridge.server`
(Ollama ignores the bridge's llama.cpp health probe, so `/health` reports the model as not ready; replies still work.)

## Evaluation
| Held-out test (300 synthetic cases) | Base | Lokol Health |
|---|---|---|
| Follows the output protocol | 0% | 99% |
| Right action (advise / refer / ask a person) | 0% | 64% |
| Cites the right manual section | 0% | 72% |
| Danger signs referred | 0% | 96% |
| Says 'ask a person' when it should | 0% | 35% |
| Replies in Pijin when asked | 73% | 100% |

Full results, safety-gate scores and the dose audit: https://github.com/VictorChenCA/lokol/blob/main/eval/results.md

## Training
LoRA (mlx-lm, all 28 layers, rank 8) on 2,704 synthetic cases generated by open-weight teachers (DeepSeek-V4.1-Flash, Kimi-K2.6 via River AI)
from the manual, then fused and quantized with llama.cpp. Data card: https://github.com/VictorChenCA/lokol/blob/main/data/DATA_CARD.md

## Other ways to run it
- Phone browser, offline after the first load: https://lokol-studio.vercel.app/demo (Add to Home screen). It adds the guideline
  lookup and the safety gate that a bare chat app lacks.
- Android app: PocketPal AI (Google Play), Models, Add from Hugging Face, search `VictorChenCA/lokol-health-qwen3-1.7b-gguf`, download `lokol-health-qwen3-1.7b-Q4_K_M.gguf`, then paste the
  system prompt from `system`. Step by step: https://github.com/VictorChenCA/lokol/blob/main/deploy/ANDROID.md
- Laptop or clinic PC, model + bridge in one command (macOS/Linux, needs llama.cpp and Python 3.10+):
  `curl -fsSL https://raw.githubusercontent.com/VictorChenCA/lokol/main/deploy/lokol-laptop.sh | bash -s -- --model 1.7b`
- By hand: `llama-server -m lokol-health-qwen3-1.7b-Q4_K_M.gguf --jinja -c 4096 --port 8080`, then the Lokol bridge (`LLM_BACKEND=llama`).

## Limitations
Synthetic training and test data; Pijin not reviewed by a native speaker; children's manual only; small models over-refer and can invent doses
(the Lokol dose guard replaces unsupported doses). Guideline excerpts are (c) Ministry of Health and Medical Services, Solomon Islands.