Llama Guard
Meta's open-weight safety classifiers that label LLM prompts and responses as safe or unsafe against the MLCommons hazard taxonomy.
Tags
Description
Llama Guard 4 is a 12B natively multimodal classifier (text plus multiple images) pruned from Llama 4 Scout, combining Llama Guard 3-8B and 3-11B-vision; it adds a Code Interpreter Abuse category for tool-call use. Weights are gated on Hugging Face under the Llama 4 Community License; earlier versions use the Llama 2/3/3.2 community licenses. Part of Meta's Purple Llama project.
Related
- Llama Prompt Guard 2: Meta's small open-weight classifiers (86M and 22M) that flag prompt injection and jailbreak attempts in user input and untrusted content.
- LlamaFirewall: Meta's open-source guardrail framework for agents that chains scanners for prompt injection, goal hijacking and insecure generated code.
- NeMo Guardrails: NVIDIA's open-source toolkit for adding programmable guardrails to LLM-based conversational systems.
- Guardrails AI: Open-source Python framework for adding input/output guardrails (validators) to LLM applications.
Sources
- https://github.com/meta-llama/PurpleLlama
- https://github.com/meta-llama/PurpleLlama/blob/main/Llama-Guard4/12B/MODEL_CARD.md
- https://huggingface.co/meta-llama/Llama-Guard-4-12B
- https://dev.meta.ai/llama/docs/model-cards-and-prompt-formats/llama-guard-4
Machine-readable
- JSON
- Markdown
- Source file on GitHub (edit via pull request)