README.md

---
license: other
license_name: license
license_link: LICENSE
configs:
- config_name: default
  data_files:
  - split: standard
    path: data/standard-*
  - split: high_quality
    path: data/high_quality-*
dataset_info:
  features:
  - name: level
    dtype: int32
  - name: language
    dtype: string
  - name: text_only
    dtype: bool
  - name: image
    dtype: image
  - name: id
    dtype: string
  - name: text
    dtype: string
  - name: label
    dtype: string
  splits:
  - name: standard
    num_bytes: 4548122261.668258
    num_examples: 108677
  - name: high_quality
    num_bytes: 1699405334.9488075
    num_examples: 40617
  download_size: 5773349370
  dataset_size: 6247527596.617065
---

# M3Kang: A Multilingual Multimodal Mathematical Reasoning Dataset from Kangaroo Problems
## Introduction

Despite state-of-the-art vision-language models (VLMs) have demonstrated strong reasoning capabilities, their performance in multilingual mathematical reasoning remains underexplored. To bridge this gap, we introduce M3Kang, the first massively multilingual, multimodal mathematical reasoning dataset for VLMs. It is derived from the Kangaroo Math Competition, the world’s largest mathematics contest, which annually engages over six million participants under the age of 18 across more than 90 countries. M3Kang includes 1,747 unique multiple-choice problems organized by grade-level difficulty, with translations into 108 culturally diverse languages. The dataset consists of two splits: standard, used in the paper for benchmarking and containing 108,677 problems, and high-quality, a smaller subset of problems with higher-quality translations, totaling 40,617 problems.

This dataset is published as part of our paper [M3Kang: Evaluating Multilingual Multimodal Mathematical Reasoning in Vision-Language Models](https://arxiv.org/abs/2601.16218).

## Description
The M3Kang dataset is built from problems in the Kangaroo Math Competition, the largest annual international contest for primary and secondary students. Problems are organized by grade level and sometimes accompanied by figures (see “Sample Images” below for an example). We sourced data from the original PDFs of the 2007–2024 editions organized by the Catalan Math Society in Catalonia, Spain, and processed it through a pipeline described below. In addition to processing and formatting the problems into a proper data set structure compatible with Hugging Face, our processing pipeline automatically translates the problems to many other languages and performs quality checks on the translations, to obtain the final multilingual data set. The final dataset includes each problem in two formats: a text version of the question and an image containing the full problem (question, multi-choice answers, and optional figure, and the correct answer) with translations into more than 100 languages.

## Sample Images
Example of problem found in the M3Kang dataset, with translations to English, Catalan, Spanish, and German:

![picture1](https://cdn-uploads.huggingface.co/production/uploads/68e7940f7b9f8592b18752b2/RPpwdPn_ydfv8saTISEbI.png)

## Dataset Details

| | |
|---|---|
| Type of problems | Math, logic, reasoning |
| Number of English problems	 | 1747 |
| Multimodal? | Yes, text + image |
| Multilingual?	| Yes, it contains translations to 108 languages |
| Number of levels | 8 levels: from school grade 5 through grade 12 |
| Total # of problems | 108,677 |

Note that because our translation pipeline includes a quality assurance stage that discards low quality translations, the number of problems in non-English languages is in general smaller than 1747. See the following table for the exact number of problems in each language for both splits. 

![m3kang_datasetoverview](https://cdn-uploads.huggingface.co/production/uploads/68e7940f7b9f8592b18752b2/7ydZqSIkObbrrDEhwSHIU.png)

## Data Format
A sample from our dataset consists in an "id", containing a unique identifier of each problem and all its translations, a string of "text", containing the problem statement, an "image", containing a snapshot of the problem (which includes the problem statement, the answer key and any accompanying figures), a "label" field, containing the correct answer, a "language" field, containing the language of the problem, a "level", containing the level (school grade) of the problem and a boolean "text_only", indicating whether the problem contains a figure. For example:

```json
{
  "id": "lvl-6_2007_10",
  "level": 6,
  "language": "eng",
  "text_only": false,
  "text": "In the figure on the right, you can
    see a triangle ABC in which two segments have
    been drawn from vertices A and B to points on
    the opposite sides, thereby dividing the triangle
    into nine non-overlapping regions. If we draw a
    total of eight segments to the opposite sides,
    four from A and four from B, into how many
    non-overlapping regions will the triangle be
    divided?",
  "label": "B",
  "image": 
```
![picture2](https://cdn-uploads.huggingface.co/production/uploads/68e7940f7b9f8592b18752b2/RDAoGgOcntNXg70E50_-5.jpeg)
```json
}
```

## Dataset Collection Process

We sourced data from the original PDFs of the 2007–2024 editions organized by the Catalan Math Society (SCM) in Catalonia, Spain. We gained access to this dataset following formal approval from the SCM, who authorized our use of the data for research purposes via an official letter. The original data was available in the form of PDF files, each file containing a test of math multi-choice Kangaroo problems (e.g., a test with 24 problems, for a given year and a given school level). The input data also included the correct answer to each problem.

We developed a pipeline to process the set of input PDFs and automatically (with some human-in-the-loop support) generate the data set in a format compatible with Hugging Face. The dataset generation pipeline relies exclusively on software rigorously reviewed and approved by our legal team. This pipeline is also described in the paper [M3Kang: Evaluating Multilingual Multimodal Mathematical Reasoning in Vision-Language Models](https://arxiv.org/abs/2601.16218).

## Dataset license
This dataset is intended for research purposes only. See [Data License Agreement - Research Use](LICENSE.pdf)

## Dataset Citation Instructions
Please cite our paper if you use this dataset in your research.

```bibtex
@misc{torrescamps2026m3kangevaluatingmultilingualmultimodal,
      title={M3Kang: Evaluating Multilingual Multimodal Mathematical Reasoning in Vision-Language Models}, 
      author={Aleix Torres-Camps and Nathaniel Mitrani Hadida and Víctor Conchello Vendrell and Àlex Batlle Casellas and Arnau Padrés Masdemont and Jordi Ros-Giralt},
      year={2026},
      eprint={2601.16218},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2601.16218}, 
}
```

### Acknowledgements
 
We gratefully acknowledge the Societat Catalana de Matemàtiques (SCM) for authorizing access to the Catalan Kangaroo problems for non-commercial use, which served as the foundation for this dataset.

### Qualcomm AI Research
At Qualcomm AI Research, we are advancing AI to make its core capabilities – perception, reasoning, and action – ubiquitous across devices. Our mission is to make breakthroughs in fundamental AI research and scale them across industries. By bringing together some of the best minds in the field, we’re pushing the boundaries of what’s possible and shaping the future of AI.
 
Qualcomm AI Research continues to invest in and support deep-learning research in computer vision. The publication of this dataset for use by the AI research community is one of our many initiatives.
Find out more about [Qualcomm AI Research](https://developer.qualcomm.com/forums/software/ai-research-datasets).
For any questions or technical support, please contact us at [research.datasets@qti.qualcomm.com](mailto:research.datasets@qti.qualcomm.com)
 
*Qualcomm AI Research is an initiative of Qualcomm Technologies, Inc.*