Transformed SAE datasets into English varieties

该数据集包含了六个基准数据集的转换版本，这些版本被转化为38种英语变体，旨在评估大型语言模型在面对英语多样性时的语言稳健性。此外，该数据集突显了在非标准英语变体上评估时，大型语言模型之间存在显著的性能差异，特别是在英语作为第二语言的变体上，准确度最多下降了46.3%。它为标准美语与语言模型性能之间的语言距离提供了深入见解。该数据集的规模涉及从6个原始数据集中衍生出的38种英语变体，其任务是对大型语言模型的的语言稳健性进行评估。

This dataset comprises adapted versions of six benchmark datasets, converted into 38 varieties of English, designed to evaluate the linguistic robustness of Large Language Models (LLMs) against English diversity. Furthermore, it reveals significant performance disparities among LLMs when assessed on non-standard English varieties; notably, accuracy decreases by up to 46.3% on English varieties used as a second language. This resource offers in-depth insights into the linguistic distance between Standard American English and LLM performance. The scale of the dataset encompasses 38 English varieties derived from the 6 original datasets, with its core objective being the evaluation of LLMs’ linguistic robustness.