ARIA-SAR Models

Three fine-tuned models for aerial search and rescue drone analysis. Built for the Gemma 4 Good Hackathon.

Models Included

combined_best.pt — YOLOv9c human detection

• Fine-tuned on HERIDAL + SARD datasets (5,123 images)
• 0.861 mAP50 accuracy
• Detects people from drone altitude

efficientnet_distress.pt — EfficientNet-B0 distress classifier

• Fine-tuned on 6,024 aerial person crops
• 92% validation accuracy
• Classes: LYING_DOWN, STATIONARY, OBSCURED

gemma-sar.gguf — Gemma 4B rescue briefing generator

• Base: google/gemma-3-4b-it
• Fine-tuned with Unsloth on 100 SAR examples
• Training loss: 2.91 → 0.54
• Quantized Q8_0 GGUF via llama.cpp
• Runs via Ollama as "aria-sar"

Usage

See the ARIA GitHub repo for full deployment instructions.

# Load Gemma into Ollama
docker cp gemma-sar.gguf ollama:/tmp/gemma-sar.gguf
docker exec ollama sh -c "echo 'FROM /tmp/gemma-sar.gguf' > /tmp/Modelfile"
docker exec ollama ollama create aria-sar -f /tmp/Modelfile