Model Overview Card for Llama Guard 3 11B Vision

Basic Information

• Model Name: Llama Guard 3 11B Vision
• Developer/Creator: Meta
• Release Date: December 6, 2023
• Version: Llama 3.2
• Model Type: Multimodal (Text and Image) Content Safety Classifier

Description

Overview

Llama Guard 3 Vision is a content safety classification model designed to safeguard Large Language Model (LLM) inputs and responses by detecting harmful multimodal prompts and text responses.

Key Features

• Detects harmful content in both text and image inputs.
• Optimized for image reasoning use cases.
• Generates text output indicating safety levels and violated content categories.
• Outperforms GPT-4o and GPT-4o mini in response classification, with lower false positive rates.

Intended Use

Designed for use cases requiring the detection of harmful content in multimodal inputs and responses, such as ensuring the safety of LLM applications.

Language Support

Primarily optimized for the English language.

Technical Details

Architecture

Llama Guard 3 Vision is a Llama-3.2-11B pretrained model, fine-tuned for content safety classification.

Training Data

The model was trained using a hybrid dataset of human-generated and synthetically generated data. This includes human-created prompts paired with corresponding images, as well as benign and violating model responses generated using in-house Llama models and jailbreaking techniques.

Data Source and Size

The dataset includes a diverse range of prompt-image pairs, labeled by humans or the Llama 3.1 405B model, and covers all hazard categories defined by MLCommons. For image data, the vision encoder rescales images into 4 chunks, each of 560x560.

Diversity and Bias

The dataset was carefully curated to encompass a diverse range of prompt-image pairs, spanning all hazard categories.

Performance Metrics

Llama Guard 3 Vision is evaluated on an internal test set following the MLCommons hazard taxonomy. Llama Guard 3 Vision demonstrates strong performance in categories such as Indiscriminate Weapons and Elections, achieving F1 scores exceeding 0.69 in every category.

Comparison to Other Models

Llama Guard 3 Vision outperforms GPT-4o and GPT-4o mini, particularly in response classification, with higher F1 scores and significantly lower false positive rates. The ambiguity of combined text and image prompts makes prompt classification more challenging compared to response classification. Llama Guard 3 Vision relies more on the model response for classification, effectively minimizing prompt-based attacks.

‍Usage

Code Samples:

The model is available on the AI/ML API platform as "Llama-Guard-3-11B-Vision-Turbo" .

API Documentation:

Detailed API Documentation is available here.

Ethical Guidelines

Llama Guard 3 Vision is fine-tuned on Llama 3.2-vision, and its performance might be limited by its (pre-)training data. It is not meant to be used as an image safety classifier nor a text-only safety classifier.

‍

Get Llama Guard 3 11B Vision Turbo API here.