KULTURE Bench
收藏资源简介:
KULTURE Bench是由延世大学开发的一个专门用于评估语言模型在韩国文化背景下理解和推理能力的基准数据集。该数据集包含3584个实例,涵盖了韩国文化新闻、成语和诗歌三个子数据集,旨在从词汇、句子和段落层面评估模型的文化理解能力。数据集的创建过程包括从韩国新闻和教科书中提取真实内容,并通过自动化和手动校对相结合的方式构建。KULTURE Bench主要用于评估语言模型在处理韩国文化相关文本时的表现,特别是在理解深层次文化背景和历史语境方面的能力。
The KULTURE Bench is a benchmark dataset developed by Yonsei University, specifically designed to evaluate the understanding and reasoning capabilities of language models within the context of Korean culture. This dataset contains 3,584 instances across three sub-datasets: Korean cultural news, idioms, and poetry, aiming to assess models' cultural comprehension abilities at the lexical, sentential, and paragraph levels. The dataset was constructed by extracting authentic content from Korean news articles and textbooks, followed by a combination of automated and manual proofreading. The KULTURE Bench is primarily used to evaluate the performance of language models when processing Korean culture-related texts, particularly their ability to understand in-depth cultural backgrounds and historical contexts.




