Janeodum/tsaro-e4b-gguf

Tsaro E4B — Multilingual Threat Extraction

Fine-tuned Gemma 4 E4B for structured threat extraction from community security reports in Hausa, Nigerian Pidgin, and English.

Part of Tsaro, a community early-warning system for northern Nigeria.

Available quantizations

File	Size	Recommended for
tsaro-e4b-q4_k_m.gguf	~5.0 GB	Primary — best quality
tsaro-e4b-q3_k_m.gguf	~4.85 GB	Minor size savings

Both require ~6-7 GB phone RAM to run.

Use with Cactus (React Native)

import { CactusLM } from 'cactus-react-native';
const lm = await CactusLM.init({
  modelUrl: 'https://huggingface.co/Janeodum/tsaro-e4b-gguf/resolve/main/tsaro-e4b-q4_k_m.gguf',
  contextSize: 2048,
});

Use with llama.cpp

./llama-cli -m tsaro-e4b-q4_k_m.gguf -p "your report here"

System prompt

You are Tsaro, a community security report analyzer for northern Nigeria.
Extract threat entities from reports in Hausa, Nigerian Pidgin, or English.
Output ONLY valid JSON with relevant fields: threat_type, vehicle_type,
vehicle_count, person_count, cattle_count, direction, location,
forest_reference, time_reference. Omit fields that are not mentioned.

Training

• Base: google/gemma-4-E4B-it
• Framework: Unsloth
• Data: 35,512 multilingual examples (Hausa, Pidgin, English)
• LoRA r=16 alpha=16, 2 epochs, lr=2e-4 cosine

License

Inherits Gemma Terms of Use: https://ai.google.dev/gemma/terms