Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversations
Introduces Llama Guard, an input-output safeguard built by instruction-tuning Llama 2-7B on a curated safety-risk taxonomy. The model classifies both user prompts and model responses as safe or unsafe and names the violated categories, and its weights were released publicly. On benchmarks including the OpenAI Moderation Evaluation dataset and ToxicChat, it matched or exceeded the content-moderation tools available at the time.
Publisher
Meta AI
Published
7 Dec 2023
Added
today
DOI
—
Key Findings
- Treats conversational safety as two separate classification jobs: prompt-harm and response-harm, each judged against an explicit category taxonomy
- Instruction-tuned from Llama 2-7B on a curated dataset, with model weights released publicly for adaptation
- Matched or exceeded contemporary moderation tools on the OpenAI Moderation Evaluation dataset and ToxicChat
Methodology Notes
Technical report (arXiv preprint) from the model's own developers, not an independent evaluation; the benchmark numbers are the authors' own runs on public moderation datasets.
Sources
arXiv preprint (primary)
Authors
Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, Madian Khabsa
Tags
Cite This
APA
Hakan Inan et al. (2023). Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversations. Meta AI. https://arxiv.org/abs/2312.06674