Open weight classifiers label prompts and responses against a risk taxonomy and can run next to your agent without sending data to a third party. Learn their categories and how to tune thresholds for your own policy.
Read in this order
- OfficialLlama Guard 3 8B model card (Meta PurpleLlama)Start here: the risk categories and prompt format of a widely used classifier.
- Open sourceGranite Guardian (IBM)An alternative open classifier with its own categories to compare.
- PaperLlama Guard: LLM-based Input-Output Safeguard for Human-AI ConversationsThe original Llama Guard paper, on taxonomy design and training.