AraTable
收藏资源简介:
AraTable是一个专为评估大型语言模型(LLMs)在阿拉伯语表格数据上的推理和理解能力而设计的全新且全面的基准。该数据集包含各种评估任务,如直接问答、事实核查和复杂推理,涉及广泛的阿拉伯语表格来源。AraTable的数据收集过程包括从不同领域收集表格数据,包括维基百科、现实世界数据和通过LLMs生成的数据。数据集创建过程遵循混合流程,其中初始内容由LLMs生成,随后由人类专家过滤和验证,以确保数据集的高质量。AraTable旨在解决当前LLMs在处理阿拉伯语表格数据时的认知挑战,并促进阿拉伯语结构化数据处理和分析的基础模型的发展。
AraTable is a novel and comprehensive benchmark designed to evaluate the reasoning and comprehension capabilities of large language models (LLMs) on Arabic tabular data. This dataset covers various evaluation tasks including direct question answering, fact-checking and complex reasoning, involving a wide range of Arabic tabular data sources. The data collection process of AraTable gathers tabular data from diverse domains, such as Wikipedia, real-world datasets and data generated by LLMs. The dataset creation follows a hybrid workflow, where the initial content is generated by LLMs and then filtered and validated by human experts to ensure high data quality. AraTable aims to address the current cognitive challenges faced by LLMs when processing Arabic tabular data, and promote the development of foundation models for Arabic structured data processing and analysis.




