| --- |
| license: mit |
| language: |
| - en |
| base_model: |
| - google/gemma-4-E2B-it |
| tags: |
| - gemma |
| - gguf |
| - multimodal |
| - vision |
| - wildlife-monitoring |
| - quantized |
| - audio |
| - text |
| --- |
| ## 🐘 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](https://www.kaggle.com/datasets/malithabandara/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](https://www.kaggle.com/competitions/gemma-4-good-hackathon/overview) using the Unsloth fine-tuning framework. |
| --- |
