Llama Guard

Open weight model that checks prompts and responses against a written policy.

About Llama Guard

Llama Guard 3 consists of a series of input and output moderation models designed to support developers in detecting various common types of violating content. The Llama Guard 3 models were built by fine-tuning Meta-Llama 3.1 and 3.2 models and optimized to support detection of the MLCommons standard hazards taxonomy.

Llama Guard is one of the components of Purple Llama, which the project describes as an umbrella project. The Llama Guard 3 models support Llama 3.2 capabilities, including seven new languages, a 128k context window, and image reasoning.

Drafted from github.com · not yet confirmed by Meta

How it works

Llama Guard is a Llama model fine tuned to read a conversation and say whether the last user or assistant turn breaks a safety policy, and which category it broke. The policy is written in plain text inside the prompt, so you can swap in your own taxonomy without retraining. Newer versions handle images and run in many languages, and there's a small 1B variant for cheap deployment.

You host it yourself (or through any inference provider that serves Llama weights). It's meant to sit in front of and behind a chat model as an input and output filter. Expect to tune the policy text and thresholds; out of the box it reflects Meta's categories.