CIDAR

CIDAR是由法赫德国王石油与矿业大学开发的第一个开放的阿拉伯语指令调整数据集，包含10,000个指令和输出对，旨在通过人工评审确保文化相关性。该数据集反映了阿拉伯地区的多样性，包括语言复杂性和丰富的文化背景，用于增强大型语言模型在阿拉伯语理解和文化上下文中的表现，旨在解决现有数据集在文化偏见和语言准确性方面的不足。CIDAR is the first open-source Arabic instruction-tuning dataset developed by King Fahd University of Petroleum and Minerals. It contains 10,000 instruction-output pairs, with cultural relevance guaranteed through human review. This dataset reflects the diversity of the Arab region, encompassing linguistic complexity and rich cultural backgrounds. It is designed to enhance the performance of large language models (LLMs) in Arabic comprehension and cultural contextual understanding, aiming to address the shortcomings of existing datasets regarding cultural bias and linguistic accuracy.