MMBias
收藏资源简介:
MMBias数据集由Janghorbani和de Melo创建,专门用于评估视觉-语言模型中的刻板印象偏差。该数据集包含3500个目标图像和350个英语短语,涵盖宗教、国籍、残疾和性取向等14个特定目标群体。数据集通过配对文本短语和相应图像,便于创建多模态嵌入,用于偏差检测实验。创建过程包括对不同目标类别的图像和文本进行精心选择和配对,旨在分析多模态模型中偏差的动态交互,特别是在公共安全和医疗等关键领域。
The MMBias dataset, created by Janghorbani and de Melo, is specifically designed to evaluate stereotype bias in vision-language models. It contains 3500 target images and 350 English phrases, covering 14 specific target groups including religion, nationality, disability, and sexual orientation. By pairing textual phrases with their corresponding images, the dataset facilitates the creation of multimodal embeddings for bias detection experiments. Its construction process involves meticulous selection and pairing of images and texts across different target categories, aiming to analyze the dynamic interactions of bias in multimodal models, particularly in critical domains such as public safety and healthcare.

- 1More is Less? A Simulation-Based Approach to Dynamic Interactions between Biases in Multimodal Models穆罕默德·本·拉希德政府学院 · 2024年



