Now available on GitHub. This is the version released with the original paper. It contains 2 million (question, answer) pairs per module, with questions limited to 160 characters in length, and answers to 30 characters in length. Note the training data for each question type is split into "train-easy", "train-medium", and "train-hard". This allows training models via a curriculum. The data can also be mixed together uniformly from these training datasets to obtain the results reported in the paper. Categories: algebra (linear equations, polynomial roots, sequences), arithmetic (pairwise operations and mixed expressions, surds), calculus (differentiation), comparison (closest numbers, pairwise comparisons, sorting), measurement (conversion, working with time), numbers (base conversion, remainders, common divisors and multiples, primality, place value, rounding numbers), polynomials (addition, simplification, composition, evaluating, expansion), and probability (sampling without replacement)

See the original paper here

Download the data set and code from GitHub here