wassname/tiny-mfv
收藏资源简介:
Tiny Moral-Foundations Vignettes是一个用于文本分类任务的数据集,专注于道德基础理论、对齐、评估和AI安全。数据集包含132个道德调查问题,来源于Clifford et al. (2015)的研究,并进行了改编以适应大型语言模型(LLMs)的使用。数据集分为三个配置:clifford(原始道德基础问题)、scifi(科幻/奇幻改编版本)和airisk(AI风险相关改编版本)。每个配置有两个分割:other_violate(第三人称原始文本)和self_violate(第一人称改写文本)。数据集的目的是通过评估LLMs在道德判断上的表现,包括错误程度和视角偏差。
Tiny Moral-Foundations Vignettes is a dataset for text-classification tasks, focusing on moral foundations theory, alignment, evaluation, and AI safety. The dataset includes 132 moral survey questions derived from Clifford et al. (2015), adapted for use with large language models (LLMs). It consists of three configurations: clifford (original moral foundations vignettes), scifi (sci-fi/fantasy adapted versions), and airisk (AI-risk related adapted versions). Each configuration has two splits: other_violate (third-person original text) and self_violate (first-person rewritten text). The dataset aims to evaluate LLMs performance in moral judgments, including wrongness and perspective bias.




