IFM/ArabicMMLU_verified

ArabicMMLU (verified)

A filtered copy of ArabicMMLU, the Arabic multiple-choice benchmark built from school exams across educational levels in several Arab countries, with test samples flagged by a multi-judge audit removed. 41 configs: one per subject/level (40) plus All, each with a test and a dev split.

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: MBZUAI/ArabicMMLU, revision 7aa530e2893ac420352b3f5c1a1310c010e9758b (files <Subject>/test.csv and <Subject>/dev.csv, plus All/test.csv and All/dev.csv).
• No local edits: every kept row has the same values as the corresponding source row, in the original order, with the original columns.
• test splits: only the excluded rows are removed. All/test is the source All/test with the same excluded rows (matched by ID) removed, so it contains exactly the union of the 40 filtered subject test splits.
• dev splits (the few-shot examples, 120 rows in All/dev) were not audited and are byte-identical to the source files.

Columns

ID, Source, Country, Group, Subject, Level, Question, Context (a reading passage, empty for most rows), Answer Key, Option 1 to Option 5 and is_few_shot (0 in test, 1 in dev), as in the source. Answer Key is the gold letter A to E, where A refers to Option 1, B to Option 2, and so on. Questions have two to five options; unused option columns are empty (in one row, ID 6249, Option 5 holds a single space). Country, Level and Context may be empty.

Audit and filtering

Samples were screened with the multi-judge audit described in IFM/discarded_samples (revision 3a3c1339, config ArabicMMLU). 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 names its subject (subset) and a row index (doc_id), which equals the 0-based row index of that subject's test.csv in the source revision listed above. Every matched row was additionally checked against the discarded record's gold label and options (all 765 match) and its question text (it equals Question, or, for the 29 rows with a reading passage, contains both Context and Question). All other rows are kept unchanged and in their original order, with the original column names and values. Only the test splits were audited.

Config	Original rows	Excluded	Kept
Total (= All)	14455	765	13690

Per-subject counts

Config	Original rows	Excluded	Kept
Accounting (University)	74	4	70
Arabic Language (General)	612	26	586
Arabic Language (Grammar)	365	32	333
Arabic Language (High School)	390	16	374
Arabic Language (Middle School)	27	1	26
Arabic Language (Primary School)	252	13	239
Biology (High School)	1409	86	1323
Civics (High School)	87	16	71
Civics (Middle School)	236	5	231
Computer Science (High School)	261	8	253
Computer Science (Middle School)	27	1	26
Computer Science (Primary School)	190	11	179
Computer Science (University)	64	4	60
Driving Test	1211	32	1179
Economics (High School)	360	4	356
Economics (Middle School)	87	7	80
Economics (University)	137	6	131
General Knowledge	864	45	819
General Knowledge (Middle School)	172	12	160
General Knowledge (Primary School)	162	18	144
Geography (High School)	1038	85	953
Geography (Middle School)	272	18	254
Geography (Primary School)	57	4	53
History (High School)	760	31	729
History (Middle School)	203	17	186
History (Primary School)	102	7	95
Islamic Studies	639	12	627
Islamic Studies (High School)	334	12	322
Islamic Studies (Middle School)	238	23	215
Islamic Studies (Primary School)	999	23	976
Law (Professional)	314	10	304
Management (University)	75	5	70
Math (Primary School)	409	19	390
Natural Science (Middle School)	242	9	233
Natural Science (Primary School)	336	10	326
Philosophy (High School)	39	3	36
Physics (High School)	255	24	231
Political Science (University)	210	13	197
Social Science (Middle School)	241	58	183
Social Science (Primary School)	705	35	670

License

The source dataset is released under CC BY-NC 4.0 (as declared on its dataset card); this filtered copy inherits that license. See the original dataset card for details.

Citation

Please cite the original authors:

@inproceedings{koto-etal-2024-arabicmmlu,
    title = "{A}rabic{MMLU}: Assessing Massive Multitask Language Understanding in {A}rabic",
    author = "Koto, Fajri and Li, Haonan and Shatnawi, Sara and Doughman, Jad and Sadallah, Abdelrahman and Alraeesi, Aisha and Almubarak, Khalid and Alyafeai, Zaid and Sengupta, Neha and Shehata, Shady and Habash, Nizar and Nakov, Preslav and Baldwin, Timothy",
    booktitle = "Findings of the Association for Computational Linguistics: ACL 2024",
    month = aug,
    year = "2024",
    address = "Bangkok, Thailand",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.findings-acl.334/",
    doi = "10.18653/v1/2024.findings-acl.334",
    pages = "5622--5640"
}

@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/},
}