OpenMath GSM8K Masked is a dataset masked for mathematical reasoning tasks, released by NVIDIA in 2024, with related paper findings at OpenMathInstruct-1: A 1.8 Million Math Instruction Tuning Dataset. It aims to assist in synthetically generating additional solution methods for the GSM8K dataset by providing masked versions of GSM8K solutions. This approach is less likely to cause reasoning inconsistencies compared to directly using the original solutions.

This dataset was used to construct the OpenMathInstruct-1 mathematical instruction fine-tuning dataset, which contains 1.8 million question-solution pairs. Its underlying generation process utilized the permissively licensed Mixtral-8x7B model.

Dataset Composition

The dataset is stored in JSON format and includes a training set with 7,473 samples:

• Question: The text content of the original GSM8K math word problems.
• Masked Reference Solution: Key parts of the original solution process are masked, typically using letter placeholders instead of numbers or intermediate calculation results, to allow the model to complete the solution steps.
• Reference Solution: The original complete solution process, used to compare against the masked version and assist in generating new solutions.
• Expected Answer: The final answer corresponding to the math problem, which can be used to verify the correctness of generated results.

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}
}