nemotron-policy-generator

Nemotron safety policies can be generated by submitting informal content-safety rules to the tool. It produces a structured package containing Markdown source-of-truth, JSON taxonomy, and inference-ready system prompts with canonical V2 mappings.

A content safety taxonomy categorizes restricted content. For Nemotron guardrails, the generator maps rough requirements to canonical V2 categories, extends them with custom rules, and outputs JSON taxonomy for runtime enforcement and training-data labeling.

Yes, you can create custom safety categories for Nemotron-Content-Safety-Reasoning-4B. The generator extends canonical V2 categories with your custom rules and documents allow-lists, calibration notes, and refusal guidance for the target model.

Yes, Nemotron policy generation supports multimodal moderation and topic-following constraints. It produces deployment-specific guidance, model-specific emit modes, and multimodal notes for both Nemotron-3-Content-Safety and Reasoning-4B models.

To export an inference-ready system prompt for runtime guardrails, input your safety requirements and target model. The generator produces system prompt artifacts with model-specific emit modes and severity metadata for immediate deployment.

The best way to structure allow-lists and refusal guidance is by mapping requirements to canonical V2 categories. The generator documents these alongside custom taxonomy extensions and calibration notes for consistent runtime enforcement across Nemotron safety models.