New dataset and models advance multimodal AI safety assessment

Researchers have introduced SafeAtlas-VL, a novel dataset and suite of guard models designed to improve multimodal safety moderation. The dataset comprises 1.5 million training instances, featuring a five-level ordered scale for image, request, and response judgments across 15 harm categories. This approach moves beyond binary safety assessments to allow for nuanced comparison of risks in multimodal interactions. An accompanying benchmark, SafeAtlas-Bench, and a series of trained guard models, including an 8B parameter model that achieves state-of-the-art performance, are also released to facilitate further research in this area. AI

IMPACT Enhances the ability to detect and compare risks in multimodal AI interactions, potentially leading to safer AI deployments.

RANK_REASON The cluster describes a new academic paper introducing a dataset and models for multimodal safety. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• Gotit.pub
• Hugging Face
• SafeAtlas-Bench
• SafeAtlas Guard
• SafeAtlas-VL
• ScienceCast

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →