CrowS-Pairs

CrowS-Pairs 有 1508 个示例，涵盖处理九种偏见的刻板印象，如种族、宗教和年龄。在 CrowS-Pairs 中，模型以两个句子呈现：一个更刻板印象，另一个更不刻板印象。这些数据侧重于对历史上处于不利地位的群体的刻板印象，并将它们与有利群体进行对比。

CrowS-Pairs consists of 1508 examples covering stereotypes addressing nine types of bias, including race, religion, and age. In CrowS-Pairs, models are presented with two sentences: one is more stereotypical, and the other is less stereotypical. This dataset focuses on stereotypes targeting historically disadvantaged groups and contrasts them with advantaged groups.

搜集汇总

背景与挑战

背景概述

CrowS-Pairs是一个用于评估语言模型社会偏见的数据集，包含1508个示例，涵盖种族、宗教、年龄等九种偏见类型。每个示例由一对句子组成，分别代表更具刻板印象和更少刻板印象的表述，旨在对比历史上处于不利和有利的群体，帮助分析模型在偏见方面的表现。