OpenMathInstruct-1 is a mathematical instruction fine-tuning dataset containing 1.8 million problem-solution pairs, released by NVIDIA in 2024. The associated paper is "OpenMathInstruct-1: A 1.8 Million Math Instruction Tuning Dataset", and the dataset aims to enhance model capabilities in mathematical reasoning and code generation through synthetic data.

The dataset includes problems from the training subsets of GSM8K and MATH, with solutions generated by the Mixtral-8x7B model combining textual reasoning and Python code interpreter execution. The dataset is divided into training and validation sets, covering the full training set ranges of GSM8K and MATH, and provides detailed field information such as expected answers, predicted answers, and error messages.

Dataset Composition

The dataset contains the following fields:

• question: The original problem from the GSM8K or MATH training set.
• generated_solution: A synthetic solution mixing textual reasoning and code blocks.
• expected_answer: The ground truth answer provided in the original dataset.
• predicted_answer: The answer extracted by the Mixtral model from the corresponding solution (extracted from \boxed{}).
• error_message: <not_executed> if no code was used; otherwise, it is empty or contains a Python exception from the corresponding code block. The string "timeout" indicates that the code block execution exceeded 10 seconds.
• is_correct: Whether the final answer was judged correct by the scoring script.
• dataset: The source dataset, either gsm8k or math.
• generation_type: either without_reference_solution or masked_reference_solution.

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}
}