Learning path · Guardrails & Safety · 80
Llama Guard
Safety classifier models (Llama Guard family) scoring inputs and outputs against policy categories for allow/block decisions.
Why it matters
- Open-weight option for on-prem moderation.
- Composable with proprietary provider safety filters.
- Category taxonomy must map to your product policies.
Key ideas
- Policy categories
- Input and output moderation
- Self-hosting
Llama Guard models classify content into harm categories—you configure which categories block, warn, or log. Useful when sending all traffic to a cloud moderator is unacceptable. Tune thresholds per locale and product surface; generic defaults over-block or under-block niche domains. Log moderation decisions for appeals and model updates. Map Llama Guard categories to product-specific playbooks so moderators know whether to block, warn, or route to humans. Ship only after eval gates pass on representative production failures.
Updated 2026-08-09 · Full learning path