Omartificial-Intelligence-Space/Arabic-Math-SFT

---
language:
- ar
license: apache-2.0
size_categories:
- 1K<n<10K
task_categories:
- visual-question-answering
- image-to-text
tags:
- math
- geometry
- arabic
- multimodal
- education
- stem
- visual-math
- sft
pretty_name: Arabic Math SFT
dataset_info:
features:
- name: image
dtype: image
- name: problem
dtype: string
- name: solution
dtype: string
splits:
- name: train
num_examples: 5000
---

<div align="center">

# Arabic Math SFT

### A Multimodal Arabic Mathematics Dataset for Supervised Fine-Tuning

[](https://opensource.org/licenses/Apache-2.0)
[](#)
[](#)
[](#)

---

*Empowering Arabic AI with visual mathematical reasoning*

</div>

## Overview

**Arabic-Math-SFT** is a curated multimodal dataset designed for supervised fine-tuning of vision-language models on **mathematical problem-solving in Arabic**. Each sample pairs a geometric or algebraic diagram with an Arabic-language problem statement and its corresponding solution.

This dataset bridges a critical gap in Arabic AI resources — while math reasoning benchmarks exist for English and Chinese, high-quality Arabic visual math datasets remain scarce.

## Dataset Structure

| Column | Type | Description |
|:---------|:---------|:----------------------------------------------|
| `image` | `Image` | Mathematical diagram (geometry, graphs, etc.) |
| `problem`| `string` | Problem statement in Arabic |
| `solution`| `string`| Answer wrapped in `<answer>` tags |

## Examples

<table>
<tr>
<td width="50%">

**Example 1**

> **المسألة:** إذا كانت زاوية ABC تساوي 25 درجة والنقاط A و B و C تقع جميعها على الدائرة O، فما قياس زاوية AOC؟
>
> **الحل:** `<answer> 50° </answer>`

</td>
<td width="50%">

**Example 2**

> **المسألة:** في الرسم المقدم، مع متوازي الأضلاع ABCD، الخط AE يقسم زاوية BAD ويقطع CD عند النقطة E. إذا كانت أطوال AD و AB هي 3.0 و 4.0 على التوالي، ما هو قياس EC؟
>
> **الحل:** `<answer> 1 </answer>`

</td>
</tr>
</table>

## Statistics

| Metric | Value |
|:------|:------|
| Total Samples | 5,000 |
| Avg Problem Length | ~127 chars |
| Min Problem Length | 5 chars |
| Max Problem Length | 476 chars |
| Samples with Images | 100% |

## Topics Covered

The dataset spans a wide range of mathematical topics typically found in middle and high school curricula:

- **Geometry** — triangles, circles, parallel lines, angles, polygons
- **Measurement** — area, perimeter, arc length, volume
- **Algebra** — solving for unknowns from geometric relationships
- **Trigonometry** — sine, cosine, tangent in geometric contexts
- **Coordinate Geometry** — points, distances, slopes

## Usage

```python
from datasets import load_dataset

ds = load_dataset("Omartificial-Intelligence-Space/Arabic-Math-SFT", split="train")

sample = ds[0]
sample["image"].show()
print(sample["problem"]) # Arabic problem text
print(sample["solution"]) # Answer
```

## Intended Use

- **SFT** for Arabic multimodal LLMs (math reasoning)
- **Benchmarking** Arabic vision-language models on STEM tasks
- **Educational tools** for Arabic-speaking students
- **Research** in cross-lingual mathematical reasoning

## Source & Construction

This dataset was constructed by sampling **5,000** examples from [`kolerk/TON-Math-SFT`](https://huggingface.co/datasets/kolerk/TON-Math-SFT) (8,031 total)

## Citation

```bibtex
@dataset{arabic_math_sft_2026,
title={Arabic-Math-SFT: A Multimodal Arabic Mathematics Dataset},
author={Omartificial Intelligence Space},
year={2026},
publisher={Hugging Face},
url={https://huggingface.co/datasets/Omartificial-Intelligence-Space/Arabic-Math-SFT}
}
```

---

<div align="center">
<b>Built with by <a href="https://huggingface.co/Omartificial-Intelligence-Space">Omartificial Intelligence Space</a></b>
</div>