ParamBench

ParamBench is a large-scale, graduate-level benchmark dataset designed to evaluate the performance of Large Language Models (LLMs) on India-centric subjects and culturally grounded knowledge. The dataset consists of 17,275 multiple-choice questions (MCQs) in Hindi, collected from Indian competitive examination papers and their corresponding answer keys. It spans 21 diverse academic subjects, including Anthropology, Sociology, History, Law, Political Science, Economics, Philosophy, and Indian Culture, providing broad coverage of humanities, social sciences, and domain-specific knowledge. Each data instance represents a single MCQ and includes a question in Hindi, four answer options (A-D), the correct answer label, and metadata such as subject, exam name, and question type. The dataset incorporates multiple question formats Normal MCQ, Assertion-Reason, Match the List, Ordering, Fill in the Blank, and Identify Incorrect Statement- enabling fine-grained evaluation of reasoning and analytical capabilities. All questions are preserved in Hindi, ensuring authentic evaluation of linguistic and cultural understanding without reliance on translation. The dataset is released as a single test split (17,275 samples) and is intended exclusively for evaluation, enabling standardized and reproducible benchmarking of LLMs across subjects and question types. Overall, ParamBench provides a comprehensive and challenging evaluation suite for measuring subject-wise knowledge, cultural awareness, and reasoning ability of modern language models in the Indian context.