PAPERSPLEASE
收藏资源简介:
PAPERSPLEASE数据集是一个包含3700个道德困境的基准数据集,旨在研究大型语言模型在优先考虑不同层次的人类需求时的决策制定。数据集的每个条目都是一个简短的故事,描述了一个寻求进入的人的情况,这些故事是使用存在、关系和成长(ERG)理论构建的,该理论将人类需求分为三个层次。数据集还包括关于个人社会身份的线索,如种族、性别和宗教,以评估模型在社会身份线索影响下的决策。该数据集可用于评估大型语言模型在道德和社会敏感情境下的决策制定和潜在的社会偏见。
The PAPERSPECIAL dataset is a benchmark collection of 3700 moral dilemmas, developed to study the decision-making behavior of large language models (LLMs) when prioritizing human needs at different hierarchical levels. Each entry in the dataset consists of a short narrative describing the circumstances of an individual seeking entry, which is constructed based on the Existence, Relatedness, and Growth (ERG) theory that categorizes human needs into three tiers. The dataset also incorporates cues related to an individual's social identities including race, gender, and religion, to evaluate the model's decision-making under the influence of such social identity cues. This dataset can be utilized to assess the decision-making capabilities and potential social biases of large language models in moral and socially sensitive scenarios.

- 1PapersPlease: A Benchmark for Evaluating Motivational Values of Large Language Models Based on ERG Theory韩国科学技术院 · 2025年



