PersonaFeedback
收藏资源简介:
PersonaFeedback是一个大型人类标注的数据集,旨在评估大型语言模型(LLM)根据预定义的用户角色和查询提供个性化响应的能力。该数据集由8298个人类标注的测试案例组成,根据用户角色的上下文复杂性和区分两个个性化响应的难度分为简单、中等和困难三个级别。PersonaFeedback通过提供一个二进制选择的评估任务,有效衡量了模型的个性化程度。数据集的创建过程包括用户角色构建、问题生成和答案生成,所有这些数据、标注协议和评估流程都将公开,以促进LLM个性化领域的研究。
PersonaFeedback is a large human-annotated dataset intended to evaluate the capacity of large language models (LLMs) to generate personalized responses based on predefined user personas and queries. This dataset consists of 8,298 human-annotated test cases, which are categorized into three levels—simple, medium, and difficult—based on the contextual complexity of the user persona and the difficulty of distinguishing between two personalized responses. PersonaFeedback effectively measures the personalization capability of models through a binary-choice evaluation task. The dataset creation process includes three core stages: user persona construction, question generation and answer generation. All relevant data, annotation protocols and evaluation workflows will be made publicly available to promote research in the field of LLM personalization.
PersonalAILab/PersonaFeedback 数据集概述
基本信息
- 语言:英语 (en)
- 许可证:Apache 2.0 (apache-2.0)
- 数据规模:1K<n<10K
任务类别
- 主要任务:文本生成 (text-generation)
相关论文
- 论文标题:PersonaFeedback: A Large-scale Human-annotated Benchmark For Personalization
- 论文链接:https://huggingface.co/papers/2506.12915

- 1PersonaFeedback: A Large-scale Human-annotated Benchmark For Personalization中国电子科技大学, 香港中文大学(深圳), 华南农业大学, OPPO · 2025年



