In this post, you learn how to prompt Amazon Nova 2 Lite for content moderation using structured and free-form approaches, grounded in the MLCommons AILuminate Assessment Standard. The prompting techniques use the AILuminate taxonomy as an example, but they work equally well with your own custom moderation policy. You can swap in your own category definitions and the prompt structure stays the same. We also benchmark the content moderation capabilities of Amazon Nova 2 Lite against several foundation models (FMs) on three public datasets.

A model’s content moderation performance is only as good as the policy that it enforces. The MLCommons AILuminate Assessment Standard v1.1 gives you a starting point with a 12-category hazard taxonomy. The standard organizes the categories into three groups (Physical, Non-Physical, and Contextual hazards), each targeting a different class of harm.

The full taxonomy includes six additional categories. See the AILuminate Assessment Standard for complete definitions.

For more on Amazon’s approach to responsible AI, see Amazon Nova and our commitment to responsible AI and the responsible use topic in the Amazon Nova User Guide.

The following diagram illustrates a typical content moderation pipeline built on Amazon Nova 2 Lite and Amazon Bedrock. Amazon Nova 2 Lite is a very low-cost multimodal model with fast inference, making it well-suited for high-throughput content moderation.