Amazon Nova 2 Lite offers zero-shot content moderation via prompting

Amazon has introduced a prompting-based content moderation workflow using its Nova 2 Lite model on Amazon Bedrock. Unlike traditional approaches that require fine-tuning with labeled training data, Nova 2 Lite lets you update moderation policies by editing the prompt itself – no model customization needed. The system supports both structured prompts (XML or JSON) and free-form prompts, and can incorporate few-shot examples to guide the model’s output format. The model is benchmarked against several foundation models using three public datasets, grounded in the MLCommons AILuminate Assessment Standard v1.1, which provides a 12-category hazard taxonomy organized into Physical, Non-Physical, and Contextual hazards.

The recommended inference configuration uses temperature 0.7 and top-p 0.9 for consistent yet diverse outputs, with the option to lower temperature to 0 for fully deterministic results. For high-throughput pipelines, disabling reasoning mode reduces latency and cost. The pipeline consists of four stages: content entry, prompt assembly (system role, policy definitions, optional few-shot examples), inference on Nova 2 Lite, and a moderation response that includes a violation flag, violated categories, and an optional explanation. This response can then trigger allow, flag, remove, or escalate actions. Amazon emphasizes that the prompting approach works equally well with custom moderation policies – you simply swap in your own category definitions while keeping the prompt structure unchanged.