SOBACO
收藏资源简介:
SOBACO数据集由东京大学的研究团队创建,旨在评估大型语言模型(LLMs)中的社会偏见和文化常识。数据集包含三个问题类别:年龄、性别和层级关系,涵盖了与日本社会和文化背景相关的主题。SOBACO数据集以统一的问答格式呈现,每个问题都包含背景信息、诱导性上下文和文化性上下文。数据集由手工模板创建,并通过众包方式进行验证,以确保模板的合理性。SOBACO数据集可用于评估LLMs在文化常识任务上的表现,以及去偏方法对其性能的影响。
The SOBACO dataset, developed by a research team at the University of Tokyo, is designed to evaluate social biases and cultural common sense in Large Language Models (LLMs). The dataset comprises three question categories: age, gender, and hierarchical relationships, with topics tied to Japanese society and cultural context. Formatted in a unified question-answering structure, each entry in the dataset includes background information, inductive context, and cultural context. The dataset was built using handcrafted templates and validated through crowdsourcing to ensure the logical soundness of these templates. This dataset can be used to assess the performance of LLMs on cultural common sense tasks, as well as the effect of debiasing methods on their performance.

- 1Bias Mitigation or Cultural Commonsense? Evaluating LLMs with a Japanese Dataset东京大学 · 2025年



