README.md

---
dataset_info:
  features:
  - name: problem
    dtype: string
  - name: level
    dtype: string
  - name: type
    dtype: string
  - name: solution
    dtype: string
  - name: solution_variants
    list: string
  splits:
  - name: train
    num_bytes: 6120153
    num_examples: 7500
  download_size: 3239633
  dataset_size: 6120153
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
language:
- en
tags:
- math
- reinforcement-learning
---


# MATH with Solution Variants

## Dataset Summary
This dataset is a processed version of [EleutherAI/hendrycks_math](https://huggingface.co/datasets/EleutherAI/hendrycks_math), which is derived from the original [MATH dataset](https://github.com/hendrycks/math) by Hendrycks et al.

The key addition in this version is the **`solution_variants`** column. This column contains semantically equivalent variations of the ground truth answer (e.g., "0.5" vs "1/2"), generated to serve as a robust reward signal for Reinforcement Learning (RL) training.

## Dataset Structure

Each example contains:
- **`problem`**: The original mathematics problem statement.
- **`solution`**: The original step-by-step solution.
- **`solution_variants`** (New): A list of valid, equivalent short answers extracted and verified with math_verify.

### Sample
```json
{
  "problem": "What is 1/4 + 1/4?",
  "solution": "... \\boxed{1/2}",
  "solution_variants": ["1/2", "0.5"]
}
```

# Citation & Attribution

If you use this dataset, please cite the original MATH paper:

```json
@article{hendrycksmath2021,
  title={Measuring Mathematical Problem Solving With the MATH Dataset},
  author={Dan Hendrycks and Collin Burns and Saurav Kadavath and Akul Arora and Steven Basart and Eric Tang and Dawn Song and Jacob Steinhardt},
  journal={NeurIPS},
  year={2021}
}
```