Gaokao-Bench

该数据集包含了2010年至2022年中国高考的多项选择题，旨在评估大型语言模型的语言理解和符号推理能力。为了增强数据集对同义转述、符号重写和结构变化的鲁棒性，对数据集进行了增强处理，重点关注真实推理而非记忆中的语法。经过增强后，该数据集总共有1,296个问题实例，其任务是进行数学推理和语言理解评估。

This dataset contains multiple-choice questions from the Chinese National College Entrance Examination (Gaokao) spanning 2010 to 2022, and is designed to evaluate the language understanding and symbolic reasoning capabilities of large language models (LLMs). To enhance the dataset's robustness against paraphrasing, symbolic rewriting and structural variations, data augmentation was conducted, with the core focus on genuine reasoning rather than memorized grammar rules. Following augmentation, the dataset comprises a total of 1,296 question instances, with its tasks targeting mathematical reasoning and language understanding assessment.