---
pretty_name: MDK12-Bench (MDK12mini)
size_categories:
- 10K<n<100K
tags:
- benchmark
- multimodal
- reasoning
- education
---

# MDK12-Bench — MDK12mini

MDK12-Bench evaluates reasoning in multimodal large language models across multiple disciplines. This dataset hosts the three cleaned **MDK12mini** archives from the [official repository](https://github.com/LanceZPF/MDK12), preserving their original contents and filenames.

- [Technical report](https://arxiv.org/abs/2504.05782)
- [Evaluation code and quick start](https://github.com/LanceZPF/MDK12/tree/main/MDK12EvalHub)

## Download

| Archive | Size (bytes) | TSV files | Rows |
| --- | ---: | ---: | ---: |
| [MDK12mini-easy.zip](https://huggingface.co/datasets/IntJudge/MDK12/resolve/main/MDK12mini-easy.zip?download=true) | 40,114,141 | 97 | 4,259 |
| [MDK12mini-medium.zip](https://huggingface.co/datasets/IntJudge/MDK12/resolve/main/MDK12mini-medium.zip?download=true) | 41,926,350 | 101 | 3,807 |
| [MDK12mini-hard.zip](https://huggingface.co/datasets/IntJudge/MDK12/resolve/main/MDK12mini-hard.zip?download=true) | 51,705,668 | 99 | 3,885 |
| **Total** | **133,746,159** | **297** | **11,951** |

Counts exclude TSV headers and macOS metadata. Each archive contains its corresponding `MDK12mini-easy/`, `MDK12mini-medium/`, or `MDK12mini-hard/` directory. The archive names retain the original subset labels; row-level `difficulty_level` values are mixed within each subset.

## Data format

TSV filenames identify the subject/stage, question type, and modality. Shared columns are:

`index`, `question`, `category`, `answer`, `knowledge`, `analysis`, `year`, `grade_level`, `difficulty_level`.

Choice-question files additionally include `A`, `B`, `C`, and `D`. Multimodal files additionally include `image`, storing Base64-encoded JPEG images; some missing images are represented as `nan`. Column sets therefore differ across files.

## Usage

Install `huggingface_hub`, then download and extract the archives:

```python
from pathlib import Path
from zipfile import ZipFile
from huggingface_hub import snapshot_download

snapshot = Path(snapshot_download(
    repo_id="IntJudge/MDK12",
    repo_type="dataset",
    allow_patterns=["MDK12mini-*.zip"],
))

output = Path("data/MDK12")
output.mkdir(parents=True, exist_ok=True)
for difficulty in ("easy", "medium", "hard"):
    with ZipFile(snapshot / f"MDK12mini-{difficulty}.zip") as archive:
        archive.extractall(output)
```

Follow the [official evaluation instructions](https://github.com/LanceZPF/MDK12/tree/main/MDK12EvalHub) to run inference and scoring.

## Citation

The following citation is reproduced from the official repository:

```bibtex
@misc{zhou2025mdk12,
      title={MDK12-Bench: A Multi-Discipline Benchmark for Evaluating Reasoning in Multimodal Large Language Models}, 
      author={Pengfei Zhou and Fanrui Zhang and Xiaopeng Peng and Zhaopan Xu and Jiaxin Ai and Yansheng Qiu and Chuanhao Li and Zhen Li and Ming Li and Yukang Feng and Jianwen Sun and Haoquan Zhang and Zizhen Li and Xiaofeng Mao and Wangbo Zhao and Kai Wang and Xiaojun Chang and Wenqi Shao and Yang You and Kaipeng Zhang},
      year={2025},
      eprint={2504.05782},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2504.05782}, 
}
```