VLGuard consists of image-text pairs categorized into training and testing sets for safety fine-tuning of Vision Large Language Models. The dataset includes metadata files in JSON format and corresponding image archives to facilitate the development of safety baselines for multimodal AI systems.
Use Cases
- Fine-tune Vision Large Language Models for safety using the image-text pairs mapped in train.json
- Evaluate model alignment and safety guardrails using the samples provided in test.json
- Develop multimodal safety benchmarks by integrating the images from train.zip with their corresponding text descriptions
Strengths
- Contains structured metadata in train.json and test.json for supervised fine-tuning
- Includes raw image data distributed across train.zip and test.zip archives
- Introduced at ICML 2024 as a standardized benchmark for VLLM safety alignment