license: cc-by-4.0
To address the scarcity of comprehensive multimodal safety evaluation benchmarks, we introduce Falcon, a large-scale dataset designed to quantify risks in multimodal large models (MLLMs). Falcon comprises 71K multimodal samples across 13 harmful categories, including illegal activities, adversarial jailbreaking prompts, and cross-modal disinformation. The dataset is structured into two functionally distinct components:
· Training Set (69.3K): Specifically curated for safety fine-tuning of MLLMs, containing diverse adversarial examples to enhance model robustness against harmful content generation.
· Test Set (1.8K): Designed as a standardized benchmark for evaluating multimodal security, featuring unseen attack scenarios to measure end-to-end model safety performance.