README.md

---
language:
- en
license: mit
size_categories:
- 1K<n<10K
task_categories:
- image-text-to-text
tags:
- math
---

# MM_Math Datasets

Paper: [A Survey of Deep Learning for Geometry Problem Solving](https://huggingface.co/papers/2507.11936)
Code: https://github.com/majianz/gps-survey

We introduce our multimodal mathematics dataset, MM-MATH,.
This dataset is collected from real middle school exams in China, and all the math problems are open-ended to evaluate the mathematical problem-solving abilities of current multimodal models. MM-MATH is annotated with fine-grained three-dimensional labels: difficulty, grade, and knowledge points. The difficulty level is determined based on the average scores of student exams, the grade labels are derived from the educational content of different grades from which the problems were collected, and the knowledge points are categorized by teachers according to the problems' content.

## MM_Math Deacription

The MM_math description contains two documents:

1. **Image.zip**: This archive includes images used in the problems.
2. **MM_Math.jsonl**: This file contains collected middle school exam questions, including the problem statement, solution process, and 3 dimension annotations.


## Data Format
All data in **MM-Math** are standardized to the following format:

```json
{
    "question": "The text of each question statement conforms to LaTeX code.",
    "file_name": "The names of the question images in the image folder.",
    "solution": "The text of each question' soluation conforms to LaTeX code.",
    "year": "The grade level annotated from each year examination.",
    "difficult": "The difficult level annotated by  examination scores.",
    "knowledge": "Each knowledge points contained in the question, which is annotated by middle school teacher."
}
```