---
license: gemma
base_model: google/gemma-4-E2B-it
base_model_relation: finetune
library_name: transformers
tags:
  - gemma-4
  - tsaro
  - threat-extraction
language:
  - ha
  - en
pipeline_tag: text-generation
---

# Tsaro Gemma 4 E2B

Fine-tuned Gemma 4 E2B 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-e2b-it`](https://huggingface.co/google/gemma-4-e2b-it)
- **Fine-tuning:** LoRA adapter trained on Tsaro threat-report data, then merged
  into the base weights
- **Role in Tsaro:** the E2B variant is the smaller of two on-device extraction
  models, used as the fallback for older or low-RAM Android devices

## Derived models

- [`Janeodum/tsaro-e2b-gguf`](https://huggingface.co/Janeodum/tsaro-e2b-gguf) — GGUF quantization for
  on-device inference via llama.cpp / llama.rn

## Training data

Fine-tuned on threat-report examples spanning Hausa, Pidgin, and English,
including examples derived from the ACLED Nigeria conflict archive 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.