Datasets:

IFM
/

GATmath_verified

size_categories:
  - 1K<n<10K
license: other
license_name: cc-by-nc-sa
license_link: https://huggingface.co/datasets/smballaa/GATmath
language:
  - ar
task_categories:
  - multiple-choice
pretty_name: GATmath (verified)
configs:
  - config_name: algebra
    data_files:
      - split: test
        path: algebra/test-00000-of-00001.parquet
  - config_name: arithmetic
    data_files:
      - split: test
        path: arithmetic/test-00000-of-00001.parquet
  - config_name: comparisons
    data_files:
      - split: test
        path: comparisons/test-00000-of-00001.parquet
  - config_name: geometry
    data_files:
      - split: test
        path: geometry/test-00000-of-00001.parquet

GATmath (verified)

A filtered copy of GATmath, the Arabic multiple-choice math benchmark derived from the General Aptitude Test (GAT), with samples flagged by a multi-judge audit removed. Four configs: algebra, arithmetic, comparisons, geometry.

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: smballaa/GATmath, revision bec17562a1fa6308e4ac713730412a52603f0b6d (files algebra.JSON, arithmetic.json, comparisons.JSON, geometry.JSON).
The copy used for evaluation (the starting point of this filtered version) equals the upstream files except that the question text (السؤال) of 3 algebra questions and 1 geometry question was edited locally; options and answers are unchanged.
Only the GATmath files are included (not the verbal GATLc files).

Columns

السؤال (question), أ, ب, ج, د (the four options), الإجابة_الصحيحة (gold letter) and النوع (type), as in the source.

Audit and filtering

Samples were screened with the multi-judge audit described in IFM/discarded_samples (revision 3a3c1339, config GATmath). 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). Excluded samples were matched to this benchmark by the row index (doc_id) assigned by lm-evaluation-harness, which equals the 0-based row index of the source file listed above; every matched row was additionally checked against the discarded record's question text and gold label. All other rows are kept in their original order, with the column names and values of the evaluation copy, which includes the locally edited question text of source rows 2 and 27 (algebra) and 3 (geometry); the other edited algebra question was excluded by the audit.

Config	Original rows	Excluded	Kept
`algebra`	2700	89	2611
`arithmetic`	2722	59	2663
`comparisons`	1225	45	1180
`geometry`	370	6	364
Total	7017	199	6818

License

The source card states CC BY-NC-SA (version not specified). This copy inherits those terms: non-commercial use with attribution, share-alike. Check the original card before reuse.

Citation

Please cite the original authors:

@article{alballaa2025gatmath,
  title   = {GATmath and GATLc: Comprehensive benchmarks for evaluating Arabic large language models},
  author  = {AlBallaa, S. and AlTwairesh, N. and AlSalman, A. and Alfarhood, S.},
  journal = {PLoS One},
  volume  = {20},
  number  = {9},
  pages   = {e0329129},
  year    = {2025},
  doi     = {10.1371/journal.pone.0329129}
}

@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/},
}