SafeAtlas-VL is a dataset of 1.5M training instances that moves beyond binary safety decisions by using a five-level ordered scale for image, request, and response judgments. It covers 15 harm categories and 55 subcategories to enable more nuanced multimodal safety moderation.
HOW THIS AFFECTS YOU
●
builderYou can implement more granular safety guardrails that distinguish between different levels of risk in multimodal inputs.
●
policyThis dataset enables more sophisticated regulatory compliance testing for generative multimodal models.