IFM/AraDiCE-Culture_verified

AraDiCE-Culture (verified)

A filtered copy of AraDiCE-Culture, the cultural split of the AraDiCE benchmark for dialectal and cultural capabilities in LLMs: three-option multiple-choice questions, written in dialectal Arabic, about the culture of six Arab countries. Samples flagged by a multi-judge audit were removed. Six configs: Egypt, Jordan, Lebanon, Palestine, Qatar, Syria (each a single test 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: QCRI/AraDiCE-Culture, revision cdd2ed80a45a1f1180dab5a371f9aeef136a7712 (files egypt/EGYPT.json, jordan/JORDAN.json, lebanon/LEBANON.json, palestine/PALESTINE.json, qatar/QATAR.json, syria/SYRIA.json). The data files are byte-identical to those of the preceding upstream commit a030244d1ab2fa547e62ecdf021843d18d52e9b5; the later commit only changed the card.
• The Jordan, Lebanon, Palestine, Qatar and Syria files are byte-identical to the source files. egypt/EGYPT.json is the source file with one row removed; the remaining 29 rows are unchanged and in their original order.
• No rows were added, edited or reordered, and no columns were added (in particular, no answer column was added).

Columns

QID (question id, e.g. Q_EG_001), Question, Option A, Option B, Option C, as in the source. There is no gold-answer column. The upstream card does not document the answer convention; the lm-evaluation-harness AraDiCE cultural tasks, which load this dataset, use the choices in the order Option A, Option B, Option C with target index 0, i.e. Option A is the correct answer for every row. Evaluations that present the options in this fixed order will therefore always have the correct answer in first position; shuffle the options (keeping track of where Option A goes) if that is a concern.

Audit and filtering

Samples were screened with the multi-judge audit described in IFM/discarded_samples (revision 3a3c1339, config AraDiCE-Culture). Three language-model judges (GPT-5.6 Sol, Gemini 3.8 Flash, Claude Opus 5) assessed candidate samples that strong baselines answered incorrectly. All six configs were in scope; the audit statistics in that repository list 180 original samples for AraDiCE-Culture, matching the source total. 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). The single excluded sample was matched to this benchmark by its subset (egypt) and row index (doc_id 22), which is the 0-based row index in egypt/EGYPT.json (QID Q_EG_023). The match was checked against the discarded record: its question text is identical and its three options are the same texts as the source row. The audited copy presented the options in a different order from the source, so the discarded record's gold letter (B) refers to that order; the option text it marks as gold is the source's Option A, consistent with the convention above. All three judges flagged the sample because two of its options state essentially the same answer. All other rows are kept unchanged and in their original order, with the original column names and values.

Config	Original rows	Excluded	Kept
Egypt	30	1	29
Jordan	30	0	30
Lebanon	30	0	30
Palestine	30	0	30
Qatar	30	0	30
Syria	30	0	30
Total	180	1	179

License

The source dataset is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0), and this filtered copy is distributed under the same license. See the original dataset card for details.

Citation

Please cite the original authors:

@inproceedings{mousi-etal-2025-aradice,
    title = "{A}ra{D}i{CE}: Benchmarks for Dialectal and Cultural Capabilities in {LLM}s",
    author = "Mousi, Basel and Durrani, Nadir and Ahmad, Fatema and Hasan, Md. Arid and Hasanain, Maram and Kabbani, Tameem and Dalvi, Fahim and Chowdhury, Shammur Absar and Alam, Firoj",
    booktitle = "Proceedings of the 31st International Conference on Computational Linguistics",
    month = jan,
    year = "2025",
    address = "Abu Dhabi, UAE",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.coling-main.283/",
    pages = "4186--4218"
}

@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/},
}