license: mit
HARDMath2 Benchmark Dataset
This repository contains a collection of mathematical benchmark problems designed for evaluating Large Language Models (LLMs) on mathematical reasoning tasks.
Building
Save .csv file exported from Google Sheet to raw_csv folder and run csv_to_yaml.py to convert all of the .csv file sto .yaml. Then push the changes to remote and the .yaml file will automatically be converted to .jsonl and pushed to an anonymized HF repository.
The .csv file should have a descriptive name for the types of problems in the file, with underscores instead of spaces.
Data Format
Each benchmark problem in the dataset is structured as a JSON object containing the following fields:
Fields
Prompt: The input string that is fed to the LLM
Solution: A LaTeX-formatted string representing the mathematical formula that solves the question posed in the prompt
Parameters: A list of independent tokens that should be treated as single variables in the LaTeX response string. These include:
- Single variables (e.g.,
$A$,$x$) - Greek letters (e.g.,
$\epsilon$) - Complex strings with subscripts (e.g.,
$\delta_{i,j}$)
Each parameter should be separated by a semicolon (;).
- Single variables (e.g.,
Example
{
"prompt": "What is the derivative of f(x) = x^2?",
"solution": "\\frac{d}{dx}(x^2) = 2x",
"parameters": "x"
}