LlamaGuard-7b

LlamaGuard-7b is Meta's first open-source LLM-based input-output safeguard model, released on December 7, 2023 as part of the Purple Llama trust-and-safety initiative. Built on the Llama 2-7B architecture and instruction-tuned on roughly 13,000 annotated examples, it classifies both user prompts and assistant responses as safe or unsafe according to a six-category harm taxonomy covering violence and hate, sexual content, criminal planning, guns and illegal weapons, regulated substances, and suicide and self-harm.

The model operates as a generative classifier: it applies a chat template to the conversation and outputs either "safe" or "unsafe" followed by violated category codes. Developers can derive binary moderation scores from the first-token probability and apply custom thresholds. Per the HuggingFace model card, LlamaGuard achieves F1 scores of 94.5% on Meta's prompt test set, 84.7% on OpenAI Moderation, 62.6% on ToxicChat, and 95.3% on response classification—outperforming OpenAI Moderation API and Google Perspective API on most reported benchmarks. It is not evaluated on standard general-knowledge leaderboards such as MMLU or GPQA.

Released under the Llama 2 Community License with weights available on HuggingFace in both native Llama and transformers formats. The model inherits Llama 2's 4,096-token context window and requires approximately 14GB VRAM for full BF16 inference. It is intended as a deployable foundation for customizable content moderation pipelines in human-AI conversation systems rather than as a general-purpose chat or reasoning model.