s-nlp/conflict_bench
收藏Hugging Face2026-02-04 更新2026-02-07 收录
下载链接：
https://hf-mirror.com/datasets/s-nlp/conflict_bench
下载链接
链接失效反馈官方服务：
资源简介：
ConflictBench是一个多语言基准数据集，用于评估大型语言模型(LLMs)中的政治和地缘偏见。它包含1900-2005年间四个国家(美国、英国、中国、苏联)之间的历史冲突事件，每个事件都有中性和极端偏见的描述，涵盖7种语言(阿拉伯语、德语、英语、法语、希伯来语、俄语、中文)。内容来自维基百科；极端偏见观点是合成的并与国家立场一致。描述由76名人类注释者标注(年龄18-63岁，教育背景从高中到博士)；注释者间一致性(Fleiss κ)为0.754。
ConflictBench is a multilingual benchmark for evaluating political and geopolitical bias in LLMs. It contains historical conflict events (1900–2005) between four countries (USA, UK, China, USSR), with neutral and ultra biased descriptions per event in 7 languages (ar, de, en, fr, he, ru, zh). Content is from Wikipedia; ultra biased viewpoints are synthetic and nation-aligned. Descriptions were labeled by 76 human annotators (ages 18–63, education from high school to doctoral); inter-annotator agreement (Fleiss’ κ) is 0.754.
提供机构：
s-nlp
