Tsaro Gemma 4 E4B
Fine-tuned Gemma 4 E4B threat extraction model for Tsaro, a shared safety
system for Northern Nigeria.
What this model does
Given an unstructured report in Hausa, Pidgin, or English, this model returns
a structured threat signal — threat type, location, perpetrator and vehicle
counts, direction of movement, time references, and a confidence score — and
judges whether the message is a genuine security report at all.
Model details
- Base model:
google/gemma-4-E4B-it
- Fine-tuning: LoRA adapter trained on Tsaro threat-report data, then merged
into the base weights
- Role in Tsaro: E4B is the primary on-device extraction model — the default
on any reasonably modern Android device. The Tsaro app loads the largest model
the hardware can run, falling back from E4B to E2B to a hosted endpoint.
Training data
Fine-tuned on 35,512 examples spanning Hausa, Pidgin, and English: 2,500
synthetic threat reports plus 33,262 examples derived from the ACLED Nigeria
conflict archive, each paired with Hausa and Pidgin translations.
Intended use and limitations
Built for community safety reporting in a specific regional context. Not a
general-purpose model. Outputs are extraction assistance, not verified
intelligence.
