Nemotron Safety Policy Generator
Generate deployable custom safety policies for NVIDIA Nemotron content-safety models.
The skill is scoped to generating Nemotron safety policies and explicitly rejects S7 exceptions for sexual content involving minors; it also documents target models, runtimes, and output flows. However, it lacks explicit user-confirmation gates, rollback procedures, general sensitive-data scrubbing rules, and dependency-integrity guidance for external models and runtimes, so points are deducted.
The workflow, templates, JSON Schema, and ambiguity handling are reasonably explicit, and an evaluation report is included. However, the source provides no executable reproduction path, committed tests covering key paths, or sufficiently concrete failure diagnostics; under static calibration the score cannot exceed 10, so points are deducted.
Activation conditions, non-fit scenarios, input modes, target models, languages, and multimodal differences are clearly described. Evidence is missing for mainland-China network reachability, external model acquisition, and non-English generation quality, while some broad trigger terms may cause false activation, so points are deducted.
The documentation is layered and includes naming, versioning, templates, a schema, evaluation files, a changelog, and traceable repository and skill license information. The benchmark itself notes missing prerequisite and limitation documentation, while maintenance ownership, release update paths, and substantive version history remain incomplete, so points are deducted.
The skill specifies directly usable Markdown, JSON, and system-prompt artifacts and supplies category fields, examples, and templates that support the core generation task. However, no independently verifiable representative outputs or end-to-end results are included; artifacts still require human review and runtime integration, so the static ceiling of 7 applies and points are deducted.
The files include a metrics-bearing NVSkills-Eval report, an evaluation dataset, a schema, and cited reference materials, providing some auditability. They do not include visible CI, independently reproducible experiment records, or corroboration from multiple evidence types, and the report cannot be independently verified in this static review, so the static ceiling of 5 applies.
- Do not treat the embedded NVSkills-Eval report as independently reproduced testing for this review; publication should include runnable CI, committed tests, and key-path reproduction records.
- Add explicit rules for scrubbing PII, credentials, and other sensitive content before processing existing policies or user input, together with storage scope, access control, and recovery procedures.
- Versions, download provenance, and mainland-China reachability for Nemotron models, vLLM/SGLang/TRTLLM/Transformers, and NeMo Guardrails are not verified and should be checked before deployment.
- The skill guides policy generation and file saving without an explicit confirmation gate; production safety, legal, medical, or financial use should add human approval and rollback controls.
What does this skill do, and when should you use it?
This skill helps users create BYO custom safety policies for NVIDIA Nemotron content-safety models. It maps rough requirements or an existing policy to the Nemotron Content Safety V2 taxonomy and can add custom categories or topic-following rules. Outputs include a Markdown policy, JSON taxonomy, and drop-in inference system prompt, with optional DOCX generation. It targets both Nemotron-Content-Safety-Reasoning-4B and multimodal Nemotron-3-Content-Safety.
It reads keywords, context, or an existing policy; determines target models, deployment pattern, locales, image input, and inference flags; maps requirements to V2 categories or custom S23+ categories; fills in definitions, scope boundaries, examples, edge cases, and runtime severity metadata; and produces Markdown, JSON, and Nemotron system-prompt artifacts. It can also produce a DOCX when the user explicitly requests a formal document, sign-off, or legal review.
- A safety engineer creating a Nemotron BYO policy for an enterprise RAG assistant.
- A product team converting requirements such as “block weapons and allow cited medical information” into a structured taxonomy.
- A multimodal application team documenting visual signals for violence, weapons, hateful symbols, and privacy categories.
- An evaluation team creating labeling or evaluation rubrics with definitions, safe examples, unsafe examples, and edge cases.
- A deployment team generating model-specific classification prompts for Nemotron inference services.
What are this skill's strengths and limitations?
- Produces Markdown, JSON taxonomy, and inference-prompt artifacts in one workflow.
- Supports V2 categories, custom categories, topic-following policies, severity metadata, and edge-case guidance.
- Distinguishes the two target models’ modalities, output values, locales, and inference options.
- Defines a non-negotiable floor against allow-listing sexual content involving minors.
- It is specialized for NVIDIA Nemotron content-safety guardrails rather than general moderation-policy generation.
- The instructions reference several files under references/ and assets/, but their contents are not included in the supplied material.
- No standalone test suite, performance measurements, or real deployment validation are documented here.
- DOCX output is conditional on an explicit formal-document, sign-off, or legal-review request.
How do you install this skill?
Install the individual skill from NVIDIA’s skills collection with the documented CLI flow:
npx skills add nvidia/skills --skill nemotron-policy-generator --yes
The supplied README does not document additional runtime packages or a separate installer for this skill.
How do you use this skill?
After installation, ask for a concrete policy-generation task, for example: “Generate a multimodal enterprise-RAG BYO policy for Nemotron-3-Content-Safety that blocks weapon-assembly images and outputs Markdown, JSON, and a system prompt.” The skill defaults to both target models unless one is explicitly named.