FreshQA
收藏资源简介:
FreshQA数据集作为一个动态的问答(QA)基准测试,包含600个英语评估样本。针对大型语言模型(LLMs)的问题根据答案的特点被分为四类:答案保持不变、答案逐渐变化、答案快速变化以及答案基于错误前提。这种评估旨在审查LLMs在回答问题时是否会产生幻觉现象,以及它们是否有能力在不被误导的情况下反驳错误的事实假设。
The FreshQA dataset serves as a dynamic question-answering (QA) benchmark, encompassing 600 English evaluation samples. The questions posed to large language models (LLMs) are categorized into four classes based on the characteristics of the answers: answers that remain unchanged, answers that gradually change, answers that rapidly change, and answers based on erroneous premises. This assessment aims to examine whether LLMs produce delusional phenomena in their responses to questions and whether they possess the ability to refute erroneous factual assumptions without being misled.
FreshLLMs数据集概述
数据集基本信息
- 名称: FreshLLMs
- 关联论文: FreshLLMs: Refreshing Large Language Models with Search Engine Augmentation
- 主要用途: 用于增强大型语言模型的搜索引擎数据
数据集组成
1. FreshQA
- 最新版本: FreshQA March 24, 2025
- 更新频率: 每周或根据请求更新
- 反馈渠道: 通过数据集电子表格评论或发送邮件至freshllms@google.com
- 历史版本: 包含2024年2月至2025年3月的多个版本
2. FreshPrompt
- 工具链接: FreshPrompt notebook
3. FreshEval
- 评估方式: 提供Relaxed和Strict两种评估模式
- 评估工具:
- 推荐模型: 建议使用
gpt-4-1106-preview进行自动评估
评估方法
- 人工评估: 推荐使用人工评估检测幻觉等问题
- 自动评估:
- 标准指标: F1/exact match或recall
- LLM自动评估: 如FactScore或FreshEval
致谢
- 感谢多位贡献者对FreshQA问题和答案的更新
- 感谢SerpApi为FreshPrompt用户提供10,000次搜索赞助
引用
bibtex @misc{vu2023freshllms, title={FreshLLMs: Refreshing Large Language Models with Search Engine Augmentation}, author={Tu Vu and Mohit Iyyer and Xuezhi Wang and Noah Constant and Jerry Wei and Jason Wei and Chris Tar and Yun-Hsuan Sung and Denny Zhou and Quoc Le and Thang Luong}, year={2023}, eprint={2310.03214}, archivePrefix={arXiv}, primaryClass={cs.CL} }




