Datasets for building safe models with refusals, content moderation, PII detection, agentic safety, and audio safety capabilities.