etri-lirs/KoTSQA-v.2.0
收藏资源简介:
KoTSQA(韩语时间敏感型问答)是一个基准测试数据集,旨在评估大型语言模型(LLM)的内置参数知识与外部检索知识之间的知识冲突问题。该数据集不受LLM数据污染的影响,专注于时间敏感的问答任务,支持对检索增强生成(RAG)等技术的客观评估。它基于韩语维基百科构建,包含多种问题类型(如单跳、多跳、错误前提),并提供了训练和测试部分,其中训练集750条(用于强化学习),测试集6,750条。数据集由韩国电子通信研究院语言智能实验室和Minds Solutions共同策划,语言为韩语,采用CC BY-SA 4.0许可证。
KoTSQA (Korean Time-Sensitive Q&A) is a benchmark that is free from LLM data contamination and can evaluate the knowledge conflict between LLMs internal knowledge and external knowledge. It is a time-sensitive question-answering dataset based on Korean Wikipedia, designed to assess issues like knowledge collision in retrieval-augmented generation (RAG). The dataset includes various question types (e.g., single-hop, multi-hop, false-premise) and is split into train (750 samples) and test (6,750 samples) parts. It was curated by ETRI LIRS and Minds Solutions, funded by a government project, and is licensed under CC BY-SA 4.0.




