ianktoo/crisis-data-v3-sft

Crisis Data v3 - Emergency Preparedness SFT Dataset

Dataset Description

Crisis Data v3 is a comprehensive supervised fine-tuning (SFT) dataset focused on emergency preparedness, disaster response, and crisis management. The dataset contains approximately 2,000 high-quality instruction-response pairs designed to train AI assistants for emergency preparedness guidance.

Key Features

• Domain: Emergency preparedness, disaster response, crisis management
• Size: ~2,000 instruction-response pairs
• Format: JSONL with instruction/input/output fields
• Source: Ready.gov disaster preparedness content
• Quality: Semantically complete chunks, validated and deduplicated
• Diversity: Multiple scenario variations (first-person, third-person, different subjects/objects)

Use Cases

• Fine-tuning conversational AI for emergency guidance
• Training crisis response chatbots
• Disaster preparedness information systems
• Safety and emergency planning assistants

Dataset Structure

Each record contains:

• instruction: User's question or request (e.g., "What should I do if there's a fire in my building?")
• input: Additional context (usually empty)
• output: AI assistant's response with actionable guidance
• metadata: Source information, chunk ID, variation method

Example

{
  "instruction": "My house might flood in the next few hours, what should I do right now?",
  "input": "",
  "output": "If flooding is imminent, take these immediate steps:\n1. Move to higher ground immediately...",
  "metadata": {
    "chunk_id": "abc123",
    "source_id": "https://www.ready.gov/floods",
    "variation_method": "llm_synthesis"
  }
}

Data Collection & Processing

Source Data

• Primary Source: Ready.gov (FEMA) disaster preparedness content
• URLs: 53 web pages covering various disaster scenarios
• Content Extraction: Trafilatura-based smart extraction for clean content

Pipeline Stages

1. Web Scraping (Trafilatura)

   • 53 URLs from ready.gov
   • Clean content extraction (29% size reduction vs. basic scraping)
   • Metadata preservation

2. Semantic Chunking

   • Per-source independent chunking (no cross-source mixing)
   • Semantic boundary detection
   • 155 chunks from 50 sources
   • 78% reduction in incomplete chunks vs. character-based chunking

3. Instruction Synthesis

   • LLM-based instruction generation (GPT-4o-mini/Claude-3.5-Sonnet)
   • Dual query types: help-seeking (81%) and information-seeking (19%)
   • Context-aware response generation
   • 153 base instruction-response pairs

4. Variation Generation

   • 12 variations per base sample
   • Diverse scenarios: different subjects, objects, perspectives, urgency levels
   • Parallel generation using OpenAI and Anthropic models
   • ~1,836 variation samples

5. Quality Filtering

   • Structure validation
   • Deduplication
   • Quality checks
   • Final dataset: ~2,000 samples

Quality Metrics

• Base Synthesis Success Rate: 98.7% (153/155 chunks)
• Query Type Balance: 81% help-seeking, 19% info-seeking
• Variation Coverage: 12x multiplier for scenario diversity
• No Cross-Source Contamination: Per-source chunking ensures topic boundaries

Query Types

Help-Seeking Queries (81%)

Questions requesting actionable guidance:

• "What should I do if..."
• "How do I handle..."
• "My [situation], what steps should I take?"

Example: "I'm stuck in my car in a snowstorm. How do I stay safe?"

Information-Seeking Queries (19%)

Questions seeking knowledge or explanations:

• "What is..."
• "What does... mean?"
• "What are the signs of..."

Example: "What is an emergency preparedness plan?"

Variation Strategies

To enhance dataset diversity, each base sample was varied using:

1. Subject Changes: "I" → "my friend", "my child", "a neighbor"
2. Object Changes: "building" → "car", "house", "apartment"
3. Urgency Levels: "might happen" → "is happening now", "just happened"
4. Perspective Shifts: First-person → third-person, singular → plural
5. Natural Rephrasing: Different wording while maintaining intent

Topics Covered

• Fire safety and home fires
• Floods and water emergencies
• Earthquakes
• Hurricanes and tropical storms
• Tornadoes
• Winter storms and extreme cold
• Extreme heat
• Power outages
• Cyberattacks
• Emergency planning and preparedness
• Evacuation procedures
• Emergency kits and supplies
• Special needs and disabilities
• Children and disasters
• Pet preparedness
• And more...

Limitations

• Domain Specificity: Focused on U.S.-based emergency preparedness (Ready.gov/FEMA)
• Geographic Bias: Content may be more applicable to U.S. disaster scenarios
• Temporal Constraints: Based on Ready.gov content as of February 2025
• LLM Generation: Variation samples are synthetically generated and may contain LLM artifacts
• No Medical Advice: Dataset does not contain medical treatment guidance

Ethical Considerations

• Emergency Context: Responses should be accurate and reliable given the critical nature of emergency situations
• Verification Recommended: Users should verify guidance with official sources
• Not a Substitute: Should complement, not replace, official emergency services and guidance
• Responsibility: Developers using this dataset should ensure deployed systems clearly indicate they are AI assistants, not official emergency services

Citation

If you use this dataset in your research or applications, please cite:

@dataset{crisis_data_v3,
  author = {ianktoo},
  title = {Crisis Data v3: Emergency Preparedness SFT Dataset},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/ianktoo/crisis-data-v3-sft}
}

License

[Specify license here - e.g., CC BY 4.0, MIT, Apache 2.0]

Acknowledgments

• Content Source: Ready.gov (FEMA) for public emergency preparedness information
• Pipeline Development: SFT Pipeline framework
• LLM Providers: OpenAI (GPT-4o-mini) and Anthropic (Claude-3.5-Sonnet) for synthesis and variation generation

Maintenance & Updates

• Version: 3.0
• Last Updated: February 2025
• Maintenance: Active
• Contact: [Your contact information]

Additional Resources

• GitHub Repository
• Pipeline Documentation
• Ready.gov Official Site