SPA-VL: 100,788 Safety Preference Samples for Vision Language Models

SPA-VL contains 100,788 samples across 6 harmfulness domains and 53 subcategories, released by researcher sqrti in mid-2024. The dataset facilitates safety preference alignment for Vision Language Models (VLMs) using multimodal image-text pairs.

Use Cases

• Safety alignment of VLMs using the 100,788 preference samples
• Benchmarking model responses across 6 harmfulness domains
• Training safety guardrail models for 53 specific subcategories of harmful content

Strengths

• 100,788 total samples
• Granular taxonomy of 53 safety subcategories
• Covers 6 distinct harmfulness domains

Limitations

• Specific column schema not provided in metadata
• Potential geographic bias toward US-centric safety standards as indicated by region tags
• Lack of explicit detail on annotation source (human vs. synthetic) in the metadata

Provenance

Source

sqrti (Arxiv 2406.12030)

Freshness

Last updated July 2024; corresponds to Arxiv publication 2406.12030.

Geography

United States