README.md

---
license: cc-by-4.0
task_categories:
- image-classification
- text-classification
- object-detection
language:
- en
tags:
- multimodal
- human-rights
- conflict-detection
- satellite-imagery
- atrocity-prevention
pretty_name: Multimodal Atrocity Identification Dataset
size_categories:
- 1M<n<10M
---

# LemkinAI Multimodal Atrocity Identification Dataset

## Dataset Overview

This multimodal dataset contains comprehensive documentation of mass atrocities and human rights violations spanning 1980-2025, with 6.8+ million anonymized incident records from 195+ countries. The dataset combines textual documentation with satellite imagery and visual evidence for AI/ML research in atrocity detection, documentation, and prevention systems.

## Dataset Structure

```
LemkinAI_Multimodal_Atrocity_Identification_Dataset/
├── README.md (this file)
├── text_dataset/
│   └── [Compressed text datasets]
└── image_dataset/
    └── [Compressed image datasets]
```

## Dataset Components

### Text Dataset
The text component includes:
- **Legal Documentation**: 5,918 international tribunal records
- **Contemporary Events**: 475,000+ filtered mass atrocity events
- **Humanitarian Violations**: 15,987 documented access violations
- **Crisis Monitoring**: Conflict monitoring and analysis data
- **Military Documentation**: 280,000+ weapons and munitions records
- **Regional Collections**: Multi-region documentation covering conflict zones

### Image Dataset
The visual component includes:
- **Satellite Imagery**: 1,521 documented sites with destruction evidence
- **Conflict Documentation**: 662 satellite-documented destruction events
- **Visual Evidence**: Comprehensive visual documentation repository
- **Geospatial Analysis**: Before/after imagery for temporal analysis

## Data Schema

### Text Records
```json
{
    "incident_id": "unique_identifier",
    "date": "ISO_8601_date",
    "location": "standardized_location",
    "latitude": "decimal_degrees_WGS84",
    "longitude": "decimal_degrees_WGS84",
    "event_type": "standardized_event_classification",
    "actors": "involved_parties",
    "fatalities": "casualty_count",
    "source": "data_provenance",
    "verification_status": "VERIFIED|REPORTED|ALLEGED",
    "severity_level": "HIGH|MEDIUM|LOW",
    "description": "incident_description"
}
```

### Image Records
```json
{
    "image_id": "unique_identifier",
    "location_coordinates": [longitude, latitude],
    "capture_date": "ISO_8601_date",
    "image_type": "SATELLITE|GROUND|AERIAL",
    "resolution": "spatial_resolution_meters",
    "damage_assessment": "destruction_analysis",
    "verification_status": "VERIFIED|UNVERIFIED",
    "associated_incident_id": "linked_text_record"
}
```

## Data Sources

### Primary Sources
The dataset aggregates information from multiple anonymized sources including:
- International legal and tribunal databases
- Conflict monitoring organizations
- Humanitarian agencies and NGOs
- Satellite imagery providers
- Open-source intelligence platforms
- Regional documentation initiatives

### Temporal Coverage
- **Start Date**: 1980
- **End Date**: 2025 (ongoing)
- **Peak Coverage**: 2010-2025 (highest documentation density)

## Quality Standards

### Data Quality Metrics
- **Completeness**: Percentage of required fields populated
- **Accuracy**: Cross-validation against multiple sources
- **Timeliness**: Incident documentation delay from occurrence
- **Geographic Precision**: Coordinate accuracy (typically <100m for satellite data)

### Verification Levels
- **VERIFIED**: Multiple source confirmation with high confidence
- **REPORTED**: Single credible source documentation
- **ALLEGED**: Unconfirmed reports requiring additional validation

## Ethical Considerations

### Privacy Protection
- Personal identifiers redacted except for public officials/perpetrators
- Victim testimonies anonymized
- Location precision reduced for sensitive sites

### Access Controls
- Sensitive imagery requires research approval
- Full dataset access restricted to verified research institutions
- Public subset available for educational use

## Usage Guidelines

### Recommended Applications
- **Atrocity Early Warning Systems**: Temporal pattern analysis
- **Geospatial Conflict Monitoring**: Satellite change detection
- **Legal Evidence Analysis**: Documentation for tribunals
- **Humanitarian Response**: Crisis mapping and needs assessment
- **Academic Research**: Conflict studies and human rights research

### Technical Requirements
- **Storage**: 50GB+ for full dataset
- **Memory**: 8GB+ RAM for processing
- **Processing**: Recommended GPU for image analysis tasks
- **Software**: Python 3.8+, pandas, geopandas for data manipulation

## Citation

```bibtex
@dataset{lemkinai_multimodal_atrocity_2025,
    title={LemkinAI Multimodal Atrocity Identification Dataset},
    author={LemkinAI},
    year={2025},
    url={https://huggingface.co/datasets/LemkinAI/Multimodal_Atrocity_Identification_Dataset},
    note={Comprehensive multimodal documentation of mass atrocities and human rights violations, 1980-2025}
}
```

### Source Attribution
This dataset aggregates information from multiple sources that have been anonymized to ensure privacy and security. Users should acknowledge the collective contribution of international legal bodies, humanitarian organizations, and monitoring agencies when using this dataset.

## License

This dataset is released under Creative Commons Attribution 4.0 International (CC BY 4.0) for research and educational purposes. Commercial use requires separate licensing agreements with original data providers.

## Contact

For dataset access, technical support, or collaboration inquiries:
- **Organization**: LemkinAI
- **Dataset Issues**: Open GitHub issue in repository
- **Research Collaboration**: Contact through Hugging Face dataset page

## Dataset Statistics

- **Total Records**: 6.8+ million incidents
- **Countries Covered**: 195+
- **Temporal Span**: 45 years (1980-2025)
- **Image Collection**: 38GB+ visual evidence
- **Data Quality**: 87% green status (ready for analysis)
- **Last Updated**: November 2025

## Changelog

### Version 1.0 (November 2025)
- Initial release with unified text and image datasets
- Standardized schema across all data sources
- Comprehensive quality assessment and validation
- BigQuery integration for large-scale analysis