Community Alignment
收藏资源简介:
社区对齐数据集(Community Alignment)是迄今为止最大且最具代表性的多语言和多轮偏好数据集,由来自五个国家的标注人员提供了近20万次比较。该数据集基于负相关抽样方法收集,旨在解决算法单一文化现象,即现有的大型语言模型(LLMs)在响应上表现出高度的同质性,无法有效学习人类偏好的多样性。数据集包括超过2500个提示-响应集,每个集至少由10个标注者标注,并且28%的标注还包含了高质量的中文解释。该数据集为改善LLMs对全球多样化人群的有效性提供了一个宝贵的资源。
Community Alignment Dataset (Community Alignment) is the largest and most representative multilingual, multi-turn preference dataset to date, curated with nearly 200,000 pairwise comparisons contributed by annotators from five countries. Collected via negative correlation sampling, this dataset aims to address the algorithmic monoculture phenomenon—where existing large language models (LLMs) exhibit high homogeneity in their responses and fail to effectively learn the diversity of human preferences. It includes over 2,500 prompt-response sets, each annotated by at least 10 annotators, and 28% of the annotations also contain high-quality Chinese explanations. This dataset provides a valuable resource for improving the effectiveness of LLMs for globally diverse populations.
Community Alignment Dataset 概述
数据集基本信息
- 名称: Community Alignment Dataset
- 许可证: Creative Commons Attribution 4.0 International License (CC-BY-4.0)
- 语言: 印地语 (hi)、英语 (en)、葡萄牙语 (pt)、意大利语 (it)、法语 (fr)
- 标签: alignment, preference, reward, llm
- 数据规模: 10K < n < 100K
数据集特点
- 大规模: 包含约200,000个LLM响应的比较,来自超过3,000名独特的标注者。
- 多语言: 包含英语、法语、意大利语、印地语和葡萄牙语的比较,其中63%的比较是非英语的。
- 提示级重叠: 2,599个提示至少包含10个标注,标注者在提示之间重叠。
- 高质量自然语言解释: 27%的提示中,标注者提供了详细的解释,说明为什么他们更喜欢某个响应。
数据集用途
- 用于对齐LLM与跨文化的人类偏好。
- 支持基于社会选择和分布方法的LLM对齐。
- 用户需在使用时实施适当的过滤和审核措施,以确保生成的输出符合其内容标准。
引用信息
-
BibTex: BibTex @article{zhang2025cultivating, title = {Cultivating Pluralism In Algorithmic Monoculture: The Community Alignment Dataset}, author = {Lily Hong Zhang and Smitha Milli and Karen Jusko and Jonathan Smith and Brandon Amos and Wassim and Bouaziz and Manon Revel and Jack Kussman and Lisa Titus and Bhaktipriya Radharapu and Jane Yu and Vidya Sarma and Kris Rose and Maximilian Nickel}, year = {2025}, journal = {arXiv preprint arXiv: 2507.09650} }
-
文本引用: Zhang, L. H., Milli, S., Jusko, K., Smith, J., Amos, B., Bouaziz, W., Revel, M., Kussmann, J., Titus, L., Radharapu, B., Yu, J., Sarma, V., Rose, K., Nickel, M. (2025). Cultivating Pluralism In Algorithmic Monoculture: The Community Alignent Dataset.
反馈与联系
- 如有任何反馈或问题,请联系: communityalignment@meta.com

- 1Cultivating Pluralism In Algorithmic Monoculture: The Community Alignment DatasetFAIR at Meta · 2025年



