Nalandadata/NalandaJEENEETBench

NalandaJEENEETBench是首个用于评估大型语言模型在印度竞争性考试（JEE Mains、JEE Advanced、NEET UG）问题上的开放基准数据集。该数据集是从Nalanda Data专有的116,000多个专家策划的考试问题中精选出来的样本，包含已验证的正确答案和逐步解答。数据集分为两个部分：基准部分（800个多选题，用于评估模型）和训练样本部分（500个多选题，包含解答，用于预览训练数据质量）。数据集平衡覆盖了物理、化学、数学和生物四个学科。该数据集填补了印度竞争性考试评估套件的空白，并展示了在7B模型上通过微调实现的显著性能提升。

NalandaJEENEETBench is the first open benchmark for evaluating large language models (LLMs) on Indian competitive exam questions (JEE Mains, JEE Advanced, NEET UG). It is a curated sample from Nalanda Datas proprietary dataset of over 116,000 expert-curated examination questions with verified correct answers and step-by-step solutions. The dataset includes two splits: a benchmark split (800 MCQs for evaluation) and a train sample split (500 MCQs with solutions for previewing training data quality). It is balanced across four subjects: Physics, Chemistry, Mathematics, and Biology. This benchmark addresses the lack of standard evaluation suites for Indian competitive exams and demonstrates significant performance improvements when fine-tuning models on this data.