---
size_categories:
- 10K<n<100K
task_categories:
- multiple-choice
pretty_name: MMLU-ProX Arabic (verified)
language:
- ar
license: mit
configs:
- config_name: ar
  data_files:
  - split: validation
    path: ar/validation-*
  - split: test
    path: ar/test-*
dataset_info:
  config_name: ar
  features:
  - name: question_id
    dtype: int64
  - name: question
    dtype: string
  - name: option_0
    dtype: string
  - name: option_1
    dtype: string
  - name: option_2
    dtype: string
  - name: option_3
    dtype: string
  - name: option_4
    dtype: string
  - name: option_5
    dtype: string
  - name: option_6
    dtype: string
  - name: option_7
    dtype: string
  - name: option_8
    dtype: string
  - name: option_9
    dtype: string
  - name: answer
    dtype: string
  - name: answer_index
    dtype: int64
  - name: cot_content
    dtype: string
  - name: category
    dtype: string
  - name: src
    dtype: string
  - name: question_id_src
    dtype: int64
  splits:
  - name: validation
    num_bytes: 85490
    num_examples: 70
  - name: test
    num_bytes: 11493382
    num_examples: 11691
  download_size: 5874885
  dataset_size: 11578872
---

# MMLU-ProX Arabic (verified)

A filtered copy of the Arabic (`ar`) subset of **MMLU-ProX**, the multilingual extension of MMLU-Pro (multiple-choice questions with up to ten options across 14 categories), with test samples flagged by a multi-judge audit removed. One config, `ar`, with a `test` split (audited and filtered) and a `validation` split (not audited, unchanged).

> **What "verified" means.** Samples were removed when all three LLM judges flagged a significant defect. This is not a human review, and only samples that baseline models answered incorrectly were audited, so some defects may remain in the kept data.

## Source
- Original dataset: [`li-lab/MMLU-ProX`](https://huggingface.co/datasets/li-lab/MMLU-ProX), revision `8e6106a6c6ce1c5027e66cc338143cf997b2aa09` (files `ar/test-00000-of-00001.parquet` and `ar/validation-00000-of-00001.parquet`; both files are byte-identical at every upstream revision since `a310c4b24c87ea26db08bb62d4cc87c0fbaf7025`).
- Only the Arabic (`ar`) config is included; the other languages of MMLU-ProX are not part of this copy.
- `test`: the source test split with 68 audited samples removed (11,759 → 11,691 rows). No other edits: the kept rows are identical to the source rows, in the original order.
- `validation`: the source validation split (70 rows, 5 per category), byte-identical to the source file. It was **not audited**.

## Columns
As in the source:
- `question`: the question text (Arabic).
- `option_0` … `option_9`: the answer options. Questions have between 3 and 10 options; unused option columns are empty (null in `test`, the string `N/A` in `validation`, as in the source), always at the end.
- `answer`: the gold option letter, `A`–`J` (`A` = `option_0`, …, `J` = `option_9`).
- `answer_index`: the 0-based index of the gold option (0–9). In the source it agrees with `answer` for every row except one test row (`question_id` 3787, `answer` = `C`, `answer_index` = 1), which is kept unchanged.
- `cot_content`: chain-of-thought exemplar text (filled in `validation`, empty in `test`).
- `category`: one of 14 categories (see the table below).
- `question_id`, `src`, `question_id_src`: the item identifier, the origin of the item (for example `ori_mmlu-professional_law`, `stemez-Physics`) and its identifier there.

## Audit and filtering

Samples of the `test` split were screened with the multi-judge audit described in [`IFM/discarded_samples`](https://huggingface.co/datasets/IFM/discarded_samples) (revision `3a3c1339`, config `MMLU-ProX-Arabic`). Three language-model judges (GPT-5.6 Sol, Gemini 3.8 Flash, Claude Opus 5) assessed candidate samples that strong baselines answered incorrectly. A sample is excluded only when **all three** judges report a *significant* issue (a problem with the question, the options or the gold label that prevents reliable answering or scoring). Each discarded record identifies its sample by category (`subset`) and a row index (`doc_id`), which is the 0-based position of the row within the source test split restricted to that category, in source order. Every one of the 68 matched rows was additionally checked against the discarded record's question text, options and gold label, and the 68 matched rows are exactly the rows missing from this copy. All other rows are kept unchanged and in their original order, with the original column names and values. The `validation` split was not part of the audit.

| Category (`test`) | Original rows | Excluded | Kept |
|---|---:|---:|---:|
| `biology` | 717 | 6 | 711 |
| `business` | 789 | 3 | 786 |
| `chemistry` | 1132 | 8 | 1124 |
| `computer science` | 410 | 4 | 406 |
| `economics` | 844 | 10 | 834 |
| `engineering` | 969 | 5 | 964 |
| `health` | 687 | 8 | 679 |
| `history` | 381 | 1 | 380 |
| `law` | 959 | 9 | 950 |
| `math` | 1351 | 5 | 1346 |
| `other` | 924 | 6 | 918 |
| `philosophy` | 499 | 2 | 497 |
| `physics` | 1299 | 0 | 1299 |
| `psychology` | 798 | 1 | 797 |
| **Total** | **11759** | **68** | **11691** |

## License
MMLU-ProX is released under the MIT License (see the [original dataset card](https://huggingface.co/datasets/li-lab/MMLU-ProX)). This copy inherits that license.

## Citation
Please cite the original authors:

```bibtex
@inproceedings{xuan-etal-2025-mmlu,
    title = "{MMLU}-{P}ro{X}: A Multilingual Benchmark for Advanced Large Language Model Evaluation",
    author = "Xuan, Weihao and Yang, Rui and Qi, Heli and Zeng, Qingcheng and Xiao, Yunze and Feng, Aosong and Liu, Dairui and Xing, Yun and Wang, Junjue and Gao, Fan and Lu, Jinghui and Jiang, Yuang and Li, Huitao and Li, Xin and Yu, Kunyu and Dong, Ruihai and Gu, Shangding and Li, Yuekang and Xie, Xiaofei and Juefei-Xu, Felix and Khomh, Foutse and Yoshie, Osamu and Chen, Qingyu and Teodoro, Douglas and Liu, Nan and Goebel, Randy and Ma, Lei and Marrese-Taylor, Edison and Lu, Shijian and Iwasawa, Yusuke and Matsuo, Yutaka and Li, Irene",
    booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.emnlp-main.79/",
    doi = "10.18653/v1/2025.emnlp-main.79",
    pages = "1513--1532"
}

@misc{jais3_2026,
  title  = {Jais 3: Introducing Arabic-centric Agentic AI},
  author = {{IFM Team, Inception42 Team, and Cerebras Team}},
  year   = {2026},
  url    = {https://ifm.ai/blog/jais3/},
}
```