Offline agricultural purchase-decision model for ADTC 2026
# FieldMind Africa
> **Before you buy the chemical, ask the laptop.**
FieldMind Africa is a reproducible ADTC 2026 agriculture submission pipeline for a small, offline language model. It is designed for extension officers, cooperatives, community centres, and agro-input shops that serve smallholder farmers where connectivity and input budgets are limited.
The product's memorable behaviour is **Spend Guard**: when symptoms are ambiguous, it asks for discriminating observations and recommends the lowest-cost safe check before suggesting a purchase. It does not pretend that a text description is a laboratory diagnosis.
## Live free cloud demo
Open **FieldMind Africa on Hugging Face Spaces**. The public Space uses the trained FieldMind Africa 1.7B Q5_K_M GGUF with llama.cpp and makes no paid API calls. Purchase-decision cards return immediately; free-form diagnosis uses the local model and is slower on the free shared CPU. Free Spaces may sleep when inactive, so allow time for a cold start.
The evidence-bounded paths for cassava mosaic, flooded/yellow maize, tomato leaf spots, and purple young maize are deterministic and source-linked in both English and Kiswahili. Warm public-API checks on 23 August returned the purple-maize and Kiswahili cassava cards in about 1.9 seconds each. Unmatched diagnosis questions use the trained GGUF and remain explicitly uncertain.
The repository also includes a **2-minute 18-second final narrated demo** built entirely with local/free tooling. It shows the trained Q5 deployment, English/Kiswahili decisions, chemical-dose guard, evidence-backed diagnosis and reproducible pipeline.
The current demo is **text-only**, not image or audio multimodal. Its local, no-API language detector supports and fully replies in the project's two verified languages: English and Kiswahili. A user can still override detection from the language menu. Users choose a diagnosis, chemical-purchase, or fertilizer-purchase flow and select the country whose pro …