Source: Llama Guard 4 model card
Hugging Face model card for Meta's 12B multimodal Llama Guard 4 safety classifier: MLCommons S1-S14 categories, output format and measured recall/FPR.
Contents
- What: model card (meta-llama/Llama-Guard-4-12B)
- Publisher: Meta
- Date: released with the Llama 4 family, April 2025 (card cites the Llama 4 licence effective 2025-04-05); read 2026-11-14
- Original: huggingface.co/meta-llama/Llama-Guard-4-12B
Summary
Llama Guard 4 is a 12B dense model pruned from Llama 4 Scout, used to classify both user prompts and model responses, with text and images (several images per prompt). It answers with safe, or unsafe followed by the violated category codes.
Key claims
- Hazard taxonomy (MLCommons-based): S1 violent crimes, S2 non-violent crimes, S3 sex-related crimes, S4 child sexual exploitation, S5 defamation, S6 specialized advice, S7 privacy, S8 intellectual property, S9 indiscriminate weapons, S10 hate, S11 suicide and self-harm, S12 sexual content, S13 elections, S14 code interpreter abuse (text only).
- Supports multiple images per prompt, unlike Llama Guard 3 11B Vision; trained mostly on 2 to 5 images.
- Response classification on Meta’s internal English set: recall 69%, false positive rate 11%, F1 61% (+8 F1 over Llama Guard 3). Multilingual: recall 43%, FPR 3%, F1 51%. Single image F1 38%, multi-image F1 52%.
- Limitations: S5, S8, S13 need current factual knowledge; behaviour beyond about three images is less tested; as an LLM it can itself be attacked by prompt injection (the card points to Prompt Guard 2 for that).