morebench
收藏资源简介:
MoReBench是一个评估语言模型在程序性和多元性道德推理方面的数据集,包含AI道德顾问或AI自主行动者面临的两难困境和专家编写的评价标准。这些困境来源于日常生活、AI风险等多个领域,并按照不同的情境类型进行分类。数据集还根据五种道德框架提供了专家评价标准。
MoReBench is a dataset for evaluating language models’ performance on procedural and pluralistic moral reasoning. It encompasses moral dilemmas confronting AI moral advisors or AI autonomous agents, along with expert-developed evaluation criteria. These dilemmas are sourced from diverse domains including daily life and AI risk scenarios, and are categorized by different scenario types. The dataset additionally provides expert evaluation criteria grounded in five distinct moral frameworks.
MoReBench 数据集概述
数据集基本信息
- 许可证:CC-BY-4.0
- 数据集名称:MoReBench
- 规模:1K<n<10K
- 配置:
- morebench(默认配置)
- morebench_theory
数据集描述
MoReBench是一个包含1000个道德困境的数据集,涉及AI道德顾问帮助人类用户决策或AI智能体自主行动的场景。每个案例包含一个困境情境和一系列情境化、专家编写的评估标准。
数据结构
主要字段
-
DILEMMA:描述涉及两个行动选择的困境场景
-
DILEMMA_SOURCE:困境来源,包括:
- daily_dilemmas
- ai_risk_dilemmas
- expert_written_ethic_bowl
- expert_written_ethic_unwrapped
- expert_written_literature
- expert_written_collab
-
DILEMMA_TYPE:困境类型,包括:
- short_case
- long_case
- expert_case
-
THEORY:理论框架(morebench配置中为"neutral")
-
RUBRIC:专家编写的评估标准字典,包含:
- rubric_dimension:五个评估维度
- Identifying:识别困境中的道德考量
- Logical Process:逻辑过程
- Clear Process:清晰过程
- Helpful Outcome:有益结果
- Harmless Outcome:无害结果
- title:情境化评估标准
- weight:重要性权重(-3到3,不含0)
- rubric_dimension:五个评估维度
-
ROLE_DOMAIN:AI角色领域,包括:
- moral_advisor:道德顾问
- moral_agent:道德智能体
-
CONTEXT:困境背景设置,涵盖多个领域
理论框架扩展
morebench_theory配置在五个规范伦理学框架下收集专家编写的评估标准:
- Kantian Deontology
- Act Utilitarianism
- Aristotelian Virtue Ethics
- Scanlonian Contractualism
- Gautheierian Contractarianism
数据加载
python from datasets import load_dataset
morebench = load_dataset("kellycyy/AIRiskDilemmas", "morebench.csv")["test"] morebench_theory = load_dataset("kellycyy/AIRiskDilemmas", "morebench_theory.csv")["test"]
引用信息
bibtex @misc{chiu2025morebenchevaluatingproceduralpluralistic, title={MoReBench: Evaluating Procedural and Pluralistic Moral Reasoning in Language Models, More than Outcomes}, author={Yu Ying Chiu and Michael S. Lee and Rachel Calcott and Brandon Handoko and Paul de Font-Reaulx and Paula Rodriguez and Chen Bo Calvin Zhang and Ziwen Han and Udari Madhushani Sehwag and Yash Maurya and Christina Q Knight and Harry R. Lloyd and Florence Bacus and Mantas Mazeika and Bing Liu and Yejin Choi and Mitchell L Gordon and Sydney Levine}, year={2025}, eprint={2510.16380}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2510.16380}, }
联系方式
- 联系人:Kelly Chiu
- 邮箱:kellycyy@uw.edu




