IFM/MedAraBench_verified

MedAraBench MCQ (verified)

A filtered version of the MedAraBench test set, the Arabic medical multiple-choice benchmark of Abu-Daoud et al. (ICLR 2026), as distributed in QIMMA's copy, with samples flagged by a multi-judge audit removed.

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.

This repository does not contain the MedAraBench questions. The source dataset is public but states no license, so we publish the filter instead: which source rows to keep, plus a script that applies it to the source.

How to use

pip install pandas pyarrow huggingface_hub
python apply_filter.py medarabench_verified

apply_filter.py downloads the source at the revision below, checks every row against the fingerprints in ids.csv, and writes the 4,031-row verified test set as one file per config, medarabench_verified/<config>/data.parquet, with the same 26 configs and columns as the source. It stops with an error if the source has changed.

Or filter it yourself:

import pandas as pd

ids = pd.read_csv("hf://datasets/IFM/MedAraBench_verified/ids.csv")
source = "hf://datasets/qimma/MCQ_MedAraBench@c62cc54f8d9d9e9e4be5aca82982b58e2da3cf0e"
verified = {}
for config, rows in ids.groupby("config", sort=False):
    df = pd.read_parquet(f"{source}/{config}/data.parquet")
    keep = rows.sort_values("row")["status"].eq("kept").to_numpy()
    verified[config] = df[keep].reset_index(drop=True)  # 26 configs, 4,031 rows in total

Files

• ids.csv: one row per source row (4,927 rows):
  • config: the source config (one of the 26 below).
  • row: the 0-based row index in that config's source file.
  • status: kept (4,031) or removed (896).
  • fingerprint: the first 16 hex characters of a SHA-256 over the row's prompt, question, choices (as a JSON list) and index, used to detect changes in the source.
• apply_filter.py: rebuilds the verified set from the source.

Source

• Original benchmark: MedAraBench, released by its authors at nyuad-cai/MedAraBench (train and test CSV files; no license stated).
• File source used here: QIMMA's copy, qimma/MCQ_MedAraBench, revision c62cc54f8d9d9e9e4be5aca82982b58e2da3cf0e, files <config>/data.parquet (26 configs, 4,927 rows). This revision is the repository's only data commit. QIMMA's copy has 4,927 rows, while the authors' test CSV has 4,959; this filter is defined only over QIMMA's copy.
• Columns of the rebuilt set, as in the source: prompt (QIMMA's evaluation prompt: an Arabic instruction followed by the question and its lettered options), question, choices (4 to 6 answer options) and index (the 0-based index of the correct option in choices).

Audit and filtering

Samples were screened with the multi-judge audit described in IFM/discarded_samples (config MedAraBench). 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's subset is the source config name in lowercase with spaces and & replaced by underscores (for example gastroenterology_hepatology), and its doc_id is the 0-based position of the sample within that config's source file; ids.csv gives the matching source row (row) for each excluded sample. All other rows are kept unchanged and in their original order.

Per-config counts (26 configs)

Config	Original rows	Excluded	Kept
Anatomy	917	87	830
Histology	214	26	188
Histology and Anatomy	39	0	39
Histology and Systemic Anatomy	9	1	8
Anesthesia	81	11	70
Biochemistry	365	62	303
Cell Biology	291	94	197
Genetics	339	163	176
Physics	699	178	521
Chemistry	82	11	71
Embryology	6	1	5
Emergency Medicine	23	2	21
Gastroenterology	29	0	29
Neurology	46	3	43
Cardiology	19	2	17
Thorasic Internal	19	2	17
Oncology	12	0	12
Chest Diseases	9	3	6
Gastroenterology & Hepatology	41	3	38
Microbiology	151	36	115
Ophthalmology	259	25	234
Pathology	54	11	43
Pharmacology	66	4	62
Physiology	667	94	573
Statistics	362	52	310
Surgery	128	25	103

	Original rows	Excluded	Kept
Total	4,927	896	4,031

License

The source dataset states no license, so this repository redistributes none of its text. The files here (ids.csv, apply_filter.py and this card) are released under CC BY 4.0. The verified set you build from the source remains subject to the source's terms.

Citation

Please cite the MedAraBench authors, and QIMMA, whose copy of the benchmark is the file source used here:

@inproceedings{abudaoud2026medarabench,
    title = {{MedAraBench}: Large-Scale {A}rabic Medical Question Answering Dataset and Benchmark},
    author = {Abu-Daoud, Mouath and Kharouf, Leen and El Hajj, Omar and El Samad, Dana and Al-Omari, Mariam and Mallat, Jihad and Saleh, Khaled and Habash, Nizar and Shamout, Farah E.},
    booktitle = {The Fourteenth International Conference on Learning Representations (ICLR)},
    year = {2026},
    url = {https://arxiv.org/abs/2602.01714}
}

@misc{alqadi2026qimma,
    title = {Are {A}rabic Benchmarks Reliable? {QIMMA}'s Quality-First Approach to {LLM} Evaluation},
    author = {AlQadi, Leen and Alzubaidi, Ahmed and Alyafeai, Mohammed and Alobeidli, Hamza and Alhammadi, Maitha and Alsuwaidi, Shaikha and Alkaabi, Omar and Boussaha, Basma El Amel and Hacid, Hakim},
    year = {2026},
    eprint = {2604.03395},
    archivePrefix = {arXiv},
    primaryClass = {cs.CL},
    url = {https://arxiv.org/abs/2604.03395}
}

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