LlamaGuard — фильтрация небезопасного контента

★ 7.7 · security

llamaguard is a Claude Code skill that integrates Meta's LlamaGuard 7–8B parameter model for input and output content moderation in LLM applications. It classifies unsafe content across six safety categories — violence and hate, sexual content, guns and illegal weapons, regulated substances, suicide and self-harm, and criminal planning — achieving 94–95% accuracy. The skill covers five practical workflows: filtering user prompts before they reach the LLM, screening model responses before delivery, high-throughput serving via vLLM (up to 100 requests/sec on a single A100), wrapping the classifier in a FastAPI REST endpoint, and connecting to NVIDIA NeMo Guardrails for rule-based pipelines. It suits teams that need a production-grade AI safety layer without maintaining hand-crafted content rules.