LingoIITGN/SangrahaTox
收藏Hugging Face2026-02-05 更新2026-02-07 收录
下载链接：
https://hf-mirror.com/datasets/LingoIITGN/SangrahaTox
下载链接
链接失效反馈官方服务：
资源简介：
SangrahaTox是一个基于文化的多模态基准数据集，旨在评估视觉-语言系统在偏见、刻板印象和安全性等10个子维度上的对齐和安全性。数据集覆盖非洲、巴西、中国、印度和日本等国家和地区，包含图像和文本提示对。每个对都由Aya-Vision-32B评分，评分内容包括文本-视觉匹配、毒性存在、严重性及文化敏感性注释。数据集包含敏感内容，严格限于研究和安全分析使用。
SangrahaTox is a culturally grounded multimodal benchmark to evaluate vision-language systems on alignment and safety across 10 sub-dimensions spanning Bias, Stereotypes, and Safety. The dataset covers regions including Africa, Brazil, China, India, and Japan, and consists of image-prompt pairs. Each pair is scored by Aya-Vision-32B, capturing textual-visual match, toxicity presence, severity, and cultural notes. The dataset contains sensitive content and is strictly for research and safety analysis.
提供机构：
LingoIITGN
