MalithaBandara/EleGuard

🐘 EleGuard: Multimodal Elephant Detection

EleGuard is a specialized, multimodal Vision-Language Model (VLM) developed for the 24/7 monitoring of elephant activity in natural habitats. By leveraging infrared (IR) imagery and bioacoustic signals, EleGuard provides a robust solution for human-elephant conflict mitigation and wildlife conservation.

Model Summary

• Project Name: EleGuard
• Base Architecture: This model is a variant based on Gemma 4 E2B.
• Modality: Multimodal (Vision + Acoustic via Spectrograms).
• Format: GGUF (Optimized for edge deployment).
• Training data: EleGuard Dataset
• Training Method: Knowledge Distillation from Gemini 3.1 Flash.

Technical Innovation: Reasoning Distillation

The core breakthrough of EleGuard is the shift from simple classification to expert reasoning. Instead of training only on labels, the model was fine-tuned on "thought blocks" generated by a Teacher model (Gemini 3.1 Flash).

For every image or audio sample, the model is trained to explain its reasoning—such as identifying thermal signatures in thick brush or frequency patterns in a rumble—before outputting a final status:

• ALERT: Elephant presence confirmed.
• SAFE: No threat detected.

Dataset Details

The model was trained on a curated dataset of 2,600 samples organized into:

• Visual Imagery: High-resolution daytime and Infrared (IR) forest captures.
• Acoustic Data: Mel Spectrograms identifying vocalizations like rumbles, roars, and trumpets.
• Paired Expert Labels: Detailed JSON reasoning files for every media asset.

Usage & Deployment

This repository contains the model weights in GGUF format, specifically optimized for edge devices (Raspberry Pi, Jetson Nano, or standard laptops) using tools like llama.cpp or Ollama.

Required Files:

1. EleGuard-gemma-4-e2b-it.GGUF (Main model weights)
2. EleGuard-gemma-4-e2b-it.mmproj.GGUF (Multimodal vision projector)

Acknowledgments & Trademarks

• Gemma is a trademark of Google LLC.
• EleGuard is a model trained on a dataset based on Gemma 4 E2B.
• This project was developed for The Gemma 4 Good Hackathon using the Unsloth fine-tuning framework.

---