SPA-VL is a specialized dataset designed to make Vision Language Models (VLMs) safer while maintaining helpfulness across diverse multimodal interactions.

• Addresses safety alignment challenges specific to VLMs that process both text and visual information
• Targets 6 key harmfulness domains to reduce potentially dangerous outputs
• Fills a critical gap in the lack of large-scale safety datasets for multimodal AI systems
• Aims to improve security by design in multimodal AI applications

This research is vital for the security community as it provides a foundation for developing safer VLMs that can resist generating harmful content while still providing valuable services across applications.

SPA-VL: A Comprehensive Safety Preference Alignment Dataset for Vision Language Model