70,000 text prompts and 68,000 manually annotated images organized into a hierarchy of 12 tasks and 44 categories. The collection covers safety domains including toxicity, fairness, bias, and privacy to evaluate text-to-image generation models.
Use Cases
- Benchmark text-to-image model safety by comparing generated outputs against the 44 safety categories
- Train safety classifiers using the 68,000 manually annotated images and their corresponding toxicity or bias labels
- Evaluate model fairness and bias by testing generation performance across the 12 task domains
Strengths
- Contains 70,000 prompts designed to trigger safety-related outputs in text-to-image models
- Includes 68,000 manually annotated images across 12 distinct safety tasks
- Features a hierarchical taxonomy consisting of 44 specific safety categories
- Covers three primary safety domains: toxicity, fairness, and privacy