wassname/tiny-mcf-vignettes

Tiny Moral-Foundations Vignettes是一个用于文本分类任务的英语数据集，包含两个配置：clifford和scifi。clifford配置包含132个来自Clifford et al. (2015)研究的小故事，覆盖了Care、Fairness、Loyalty、Authority、Sanctity、Liberty等道德基础，以及一个Social Norms的负面控制。scifi配置包含51个手工编写的科幻/奇幻故事，同样覆盖了上述7个道德基础。每个小故事有四种条件：other_violate（原始第三人称违反）、other_uphold（LLM重写的第三人称支持）、self_violate（LLM重写的第一人称违反）和self_uphold（LLM重写的第一人称支持）。数据集用于快速内部循环的道德基础探测，以指导LLM检查点的调整。

Tiny Moral-Foundations Vignettes is an English dataset for text classification tasks, containing two configurations: clifford and scifi. The clifford configuration includes 132 vignettes from Clifford et al. (2015), covering moral foundations such as Care, Fairness, Loyalty, Authority, Sanctity, Liberty, plus a Social Norms negative control. The scifi configuration includes 51 hand-written sci-fi/fantasy vignettes covering the same 7 foundations. Each vignette has four conditions: other_violate (original third-person violation), other_uphold (LLM-rewritten third-person upholding), self_violate (LLM-rewritten first-person violation), and self_uphold (LLM-rewritten first-person upholding). The dataset is used for fast inner-loop moral-foundations probing to steer LLM checkpoints.