On-device clinical coaching model for community health workers (CHWs / relais communautaires) in Niger, used by the offline CommCare AI Coach. Distilled from Claude Opus 4.8 as teacher (answer-relabel SFT) on a French iCCM/CHW corpus, then quantized to Q4_K_M GGUF (2.33 GB) for CPU inference on Android via llama.cpp.

Base: Qwen3-4B
Quant: Q4_K_M (re-quantized locally; Q4_0 broke tokenizer merges)
Deploy mode: NO-THINK (/no_think + empty <think></think> prefill) for low latency
Eval (n=79 clinical, Claude-judged): clinical 2.51 / safety 2.80 / relevance 3.15 with the v2 reconciled RAG KB — the best decider config in the project's eval arc.
Role in the app: the LLM fallback behind an on-device deterministic clinical decision engine (verified condition cards); the engine handles routed clinical queries with zero inference, this model answers everything else.

Usage

bash

# llama.cpp (no-think prompt)
./llama-cli -m Qwen3-4B-CHW-Coach-v8-Q4_K_M.gguf -p "<|im_start|>user\nQ<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"

In the CommCare AI Coach the app downloads this file automatically (model id qwen3-4b-chw-v8) to the device; it can also be side-loaded over USB.

⚠️ Status

PROVISIONAL — not clinically signed off. Outputs are not a substitute for protocol or clinician judgment. The associated decision-engine cards and answer key match an internally-reconciled gold, not a clinician-verified national protocol; clinician review is pending. Use for research / development only.

idiene/Qwen3-4B-CHW-Coach-v8-GGUF

Qwen3-4B CHW Coach v8 (GGUF, Q4_K_M)

Usage

⚠️ Status

idiene/
Qwen3-4B-CHW-Coach-v8-GGUF