FloodSaathi: Offline-First Small Language Model (SLM) for Flood Safety & Early Warning

FloodSaathi is an offline-first disaster early-warning model fine-tuned from Qwen/Qwen2.5-0.5B-Instruct for India (pilot: Punjab). It provides concise, safety-verified protocols for drinking water purification, electrical wire hazards, road crossings, snakebite first aid, evacuation prep, and livestock safety in English, Hindi, and Punjabi.

Key Safety Invariants

• Grounded in verified NDMA, Punjab SDMA, and CWC flood safety protocols.
• Emergency helplines (112, 1078, 1070, 1077) are cited with zero hallucination.
• Never gives false safety guarantees; provides honest empirical guidance.

Model Details

• Architecture: Qwen2.5-0.5B-Instruct with LoRA adapters (Rank 16, Alpha 32)
• Parameters: 494M base + 18.4M trainable adapter parameters (~35 MB safetensors)
• Target Hardware: Runs on mobile devices, low-power edge devices, or offline laptops.