vekarika/farmhealth-ai
Domaine:
agriculturenatural language processingType de record:
softwaremodel
Créateur:
vek
Hôte:
Offline Agricultural & Livestock Health Advisor for African Farmers and Extension Officers
# FarmHealth AI — ADTC 2026 Laptop LLM Submission (Agriculture)
**Offline agricultural & livestock-health advisor for farmers and extension officers in
northern Nigeria** (Nasarawa, Kaduna, Adamawa, Kwali) — runs entirely on an 8 GB laptop with
zero connectivity, in **English and Hausa**, through llama.cpp on a quantized GGUF model.
Built to the official ADTC 2026 submission template, and organised around how the challenge
**actually scores** submissions.
---
## How scoring drives the design
The profiler runs the **bare GGUF model** (no app/RAG in the scored loop):
```
S_total = 0.50·S_acc + 0.30·S_perf + 0.20·S_eff − P_thermal (+ up to 10 African-use-case points)
(lm-eval MCQ) (llama-bench) (peak RSS) (>85°C)
```
- **Accuracy (50%)** = multiple-choice (lm-eval `acc_norm`) on the bare model → won by a
**fine-tuned small model** + a strong domain dataset. RAG is *not* in this loop.
- **Speed (30%) + Efficiency (20%)** → reward a **small, lean** model (target ≥15 t/s, low RAM, <7 GB hard ceiling).
- **Cross-disciplinary integration** = location-aware **offline RAG** (information retrieval) →
the qualitative score, the demo, and the live defense.
- **African use-case bonus** (≤10 pts) → English-primary model with **Hausa** support + genuine local grounding.
Architecture principle: **the model reasons; retrieval supplies what changes by place and week.**
---
## Repo map
```
.
├── metadata.json # ADTC submission metadata (schema-valid; fill team_id + github_handle)
├── download_model.sh # pulls the public GGUF into model/ (edit URL after hosting weights)
├── REPORT.md # technical writeup (problem, design, constraints, benchmarks)
├── model/ # GGUF lands here at eval time (gitignored — never commit weights)
│
├── ADTC_MASTER_SPEC.md # authoritative rules/timeline/scoring (single source of truth)
├── DOMAIN_MAP.md # knowledge coverage map (stable→fine-tune vs volatile→RAG)
├── DATASET_SPEC.m …