DiSCo2026/DiSCo_Dataset_and_Benchmark

DiSCo（分布优先的引导和上下文评估）是一个文化偏好偏见数据集和基准，用于评估和审计大型语言模型在不同文化背景下的表现。数据集包含150,816行数据，每行提供一个国家中立的生活方式问题，有四个答案选项，每个选项都配有一行文化事实字符串。所有四个选项都是同等有效且基于文化的，没有单一正确答案。数据集覆盖12个文化区域（如阿尔及利亚、中国、美国等）和3个生活方式类别（食物、家庭、休闲）。数据集结构包括问题ID、问题文本、选项、文化事实字符串等字段。支持四层上下文升级协议，用于分离默认文化先验和上下文驱动的适应性。数据集设计用于审计、基准测试和测量12个文化区域的可控性和先验粘性。

DiSCo (Distribution-first Steering and Context evaluation) is a cultural preference bias dataset and benchmark for evaluating and auditing large language models across diverse cultural contexts. The dataset consists of 150,816 rows, each presenting a country-neutral lifestyle question with four answer options. Each option is paired with a one-line cultural fact string, and all four options per item are equally valid and culturally grounded—there is no single correct answer. The dataset covers 12 cultural regions (e.g., Algeria, China, US) and 3 lifestyle categories (Food, Family, Leisure). The dataset structure includes fields such as question ID, question text, options, and cultural fact strings. It supports a four-level context escalation protocol to separate default cultural priors from context-driven adaptation. The dataset is designed for auditing, benchmarking, and measuring controllability and prior stickiness across 12 cultural regions.