SafeSteerDataset: 2,300 Contrastive Prompt Pairs for T2I Safety

NASK-PIB released the SafeSteerDataset in March 2026, providing 2,300 contrastive prompt pairs for Text-to-Image safety research. The data is organized into a granular taxonomy of 23 subcategories to isolate the geometric direction of toxicity with minimal semantic variance.

Use Cases

• Steering Text-to-Image models using the safe and unsafe prompt pairs
• Evaluating model safety across 23 specific toxicity subcategories
• Analyzing the geometric direction of toxicity in latent spaces

Strengths

• 2,300 contrastive prompt pairs
• 23 granular subcategories of toxicity
• Minimal semantic difference between safe and unsafe pairs

Limitations

• Small sample size of 2,300 records
• Limited to Text-to-Image prompt domain

Provenance

Source

NASK-PIB

Collection Method

filtered for minimal semantic difference

Freshness

Last updated March 2026.