MMSafeAware
收藏资源简介:
MMSafeAware是一个全面的多模态安全意识基准,由腾讯AI实验室推出,包含29个安全场景下的1500个经过精心策划的图像-提示对。该数据集旨在评估多模态大型语言模型在识别多模态内容安全性方面的能力,分为不安全子集和过安全子集,以评估模型正确识别不安全内容和避免过度敏感的能力。数据集经过人工标注确保质量,是目前为止最全面的多模态安全基准。
MMSafeAware is a comprehensive multimodal safety awareness benchmark launched by Tencent AI Lab, which contains 1,500 meticulously curated image-prompt pairs across 29 safety scenarios. This benchmark is designed to evaluate the capability of multimodal large language models (LLMs) in identifying the safety of multimodal content. It is divided into an unsafe subset and an over-sensitive subset, aiming to assess a model's ability to correctly identify unsafe content and avoid over-sensitivity. The dataset has been manually annotated to ensure its quality, making it the most comprehensive multimodal safety benchmark to date.

- 1Can't See the Forest for the Trees: Benchmarking Multimodal Safety Awareness for Multimodal LLMs香港科技大学, 腾讯AI实验室, 香港中文大学(深圳), 香港中文大学 · 2025年



