MathEval

MathEval is a benchmark dedicated to a comprehensive evaluation of the mathematical capabilities of large models. It encompasses over 20 evaluation datasets across various mathematical domains, with over 30,000 math problems. The goal is to thoroughly evaluate the performance of large models in tackling problems spanning a wide range of difficulty levels and diverse mathematical subfields (i.e. arithmetic, elementary mathematics, middle and high school competition topics, advanced mathematical, etc.). It serves as a trustworthy reference for cross-model comparisons of mathematical abilities among large models at the current stage and guides how to further enhance the mathematical capabilities of these models in the future.