●researcherYou can now train and evaluate guard models on fine-grained, multi-level safety judgments rather than simple binary classifications.
●policyThis allows for more nuanced safety auditing of multimodal models across specific harm subcategories.