Llama Guard

Llama Guard is a family of AI content moderation models for developers who need to check what users submit to an LLM and what the model returns. It supports text and image checks for applications that accept more than written prompts.

• Input and output checks: Evaluate user content and generated responses for common types of policy violations.
• Text and image moderation: Check written content and use image reasoning to assess visual inputs.
• Multilingual coverage: Moderate content across supported languages for applications serving users in different languages.
• Standard risk categories: Classify content using the MLCommons hazard taxonomy.
• Cyberattack content checks: Detect responses that assist cyberattacks and malicious code output intended for execution in environments with code interpreters.

The models are available through Hugging Face and build on fine-tuned Meta-Llama 3.1 and 3.2 models. Llama Guard belongs to the Purple Llama safeguards collection; Prompt Guard and Code Shield are separate tools for prompt attacks and insecure generated code.