0xKitkat/safestep-disaster-instruction

SAFEstep expanded training system

This package contains the combined, source-grounded instruction-tuning data for SAFEstep. It covers natural disasters, emergency sheltering, CPR/AED, first aid, poisoning and region-sensitive biological hazards such as snakebite and unknown mushroom ingestion.

Validated composition

• —1,004 total examples
• —780 training examples
• —224 held-out validation examples
• —16 canonical hazard categories
• —72 bundled offline playbooks
• —English, Spanish and Vietnamese prompts
• —Text, voice-transcript and image-context inputs
• —Adversarial refusal, uncertainty and accessibility cases
• —Exact iOS output contract: immediate | high | caution

Run the complete build and validation sequence:

python3 merge_training.py
python3 build_playbooks.py
python3 validate_expanded.py
python3 train_qlora.py --validate-only

The components directory preserves the independently authored base, disaster and first-aid datasets together with their source registers and coverage notes. expanded_playbooks.json is copied into the iPhone app bundle and provides deterministic offline retrieval and fallback instructions.

The image-context records teach grounded response behavior but do not visually fine-tune Gemma's image encoder. Actual snake, mushroom and wildlife detection has a separate computer-vision package in SAFEstepVisionTraining.

Safety boundary

This is a hackathon-scale research dataset—not certified medical advice, clinical training or a public warning service. Qualified emergency-management, clinical and native-language review is required before production use. Fine-tuning must never replace bundled retrieval, output validation, deterministic fallbacks or live emergency services.