OpenMath MATH Masked is a dataset released by NVIDIA in 2024, focusing on mathematical reasoning and synthetic data generation. The associated paper is "OpenMathInstruct-1: A 1.8 Million Math Instruction Tuning Dataset", which aims to generate more consistent and higher-quality synthetic mathematical data by providing masked versions of solution steps, thereby improving the performance of instruction-tuned models.

This dataset is built upon the solutions from the MATH dataset, using masking techniques to reduce reasoning inconsistencies that may arise from directly copying original solutions. It is suitable for instruction fine-tuning and synthetic data augmentation in the mathematics domain.

Dataset Composition

The dataset primarily contains the following fields:

• question: The mathematical problem.
• masked_reference_solution: The reference solution process with masking applied.
• reference_solution: The original reference solution process.
• expected_answer: The standard answer corresponding to the problem.
• level, type: Information such as difficulty level and category, subject to the actual data files.

Citation

@article{toshniwal2024openmath,
  title   = {OpenMathInstruct-1: A 1.8 Million Math Instruction Tuning Dataset},
  author  = {Shubham Toshniwal and Ivan Moshkov and Sean Narenthiran and Daria Gitman and Fei Jia and Igor Gitman},
  year    = {2024},
  journal = {arXiv preprint arXiv: Arxiv-2402.10176}
}