Hugging Face's logo

Datasets:
hamzaaouadi
/
MoroccanMedMCQA-FR

Tasks:
Question Answering
Modalities:
Text
Formats:
json
Sub-tasks:
multiple-choice-qa
Languages:
French
Size:
1K - 10K
Tags:
medical
french
morocco
benchmark
mcq
multiple-choice
DOI:
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

MoroccanMedMCQA-FR is available exclusively for non-commercial academic research purposes. By requesting access you agree to the MoroccanMedMCQA-FR Data Usage Agreement, which prohibits redistribution, commercial use, and use for training commercial AI systems. Recipients must cite the original paper in any resulting publication. Requests are reviewed manually within 14 business days. For questions contact: hamza.aouadi@usmba.ac.ma

Log in or Sign Up to review the conditions and access this dataset content.

MoroccanMedMCQA-FR: A French-Language Moroccan Medical Multiple-Choice QA Benchmark

Language: French Questions: 6,771 Specialties: 31 License: CC-BY-NC-4.0 Task: Medical MCQ Question Types: MCQU | MCQM Cognitive Focus: Understanding | Reasoning

Dataset Description

MoroccanMedMCQA-FR is the first French-language medical multiple-choice question answering (MCQ) benchmark grounded in the Moroccan medical faculty curriculum. It comprises 6,771 officially sourced MCQs drawn from past examinations of the Faculty of Medicine and Pharmacy of Fès (FMPF), Sidi Mohammed Ben Abdellah University, Morocco, covering the full 5-year medical curriculum across 31 medical specialties and spanning exam sessions from 2015 to 2023.

Unlike most existing medical QA benchmarks — which focus on English-language, US/UK-centric licensing exams with single-answer questions — MoroccanMedMCQA-FR features 77.9% multi-answer questions (where examinees must select all correct options), official past exam questions with authoritative corrections, and rich metadata including specialty, curriculum year, and exam session.

  • Paper: [PLACEHOLDER — Title TBD] (link will be added upon publication)
  • Authors:
    • AOUADI Hamza — Sidi Mohammed Ben Abdellah University, Fès, Morocco
    • NFAOUI El Habib — Sidi Mohammed Ben Abdellah University, Fès, Morocco
    • ELGAROUANI Said — Sidi Mohammed Ben Abdellah University, Fès, Morocco
  • Contact: hamza.aouadi@usmba.ac.ma
  • License: CC BY-NC 4.0

Key Features

Feature MoroccanMedMCQA-FR FrenchMedMCQA Typical English Benchmarks
Language French French English
Region Morocco / North Africa France US / UK
Medical domain General medicine (31 specialties) Pharmacy only General medicine
Questions 6,771 3,105 Varies
Answer format 77.9% multi-answer 65.2% multi-answer Varies
Options per question 4–5 (A–E) 5 (A–E) 4–5
Vocabulary size 29,284 13,000 Varies
Curriculum scope Full 5-year medical journey Pharmacy diploma Varies
Source Official past exams Official past exams Varies
Ground truth Official corrections Official corrections Varies
Clinical context field Yes (13.5%) No Varies
Split ratio 70/10/20 70/10/20 Varies
Publicly available Gated ✅ Fully open Varies

Dataset Statistics

MoroccanMedMCQA-FR comprises 6,771 questions drawn from official past exams of the Faculty of Medicine and Pharmacy of Fès, spanning 9 years of exam sessions (2015–2023) across 31 medical specialties and 5 curriculum years. The dataset is predominantly multi-answer: 77.9% of questions require selecting multiple correct options, with an average of 2.58 correct answers per question — making it significantly more challenging than most existing medical MCQ benchmarks. The distribution of correct answers per question is as follows: 1 correct option (21.6%), 2 (26.1%), 3 (29.2%), 4 (16.6%), 5 (6.0%), and questions with no answers (0.5%). A total of 914 questions (13.5%) include a clinical context vignette, with an average context length of 102.7 words (ranging from 11 to 1,049 words). The average question length is 11.9 words and the average answer option length is 6.6 words. The dataset vocabulary spans 29,284 unique words. Most questions offer five answer options (91.5%), while 8.5% have four options. Exam sessions are split between Normale (57.5%), Rattrapage (42.2%), and Exceptionnelle (0.3%) sessions.

Metric Value
Total questions 6,771
Medical specialties 31
Exam years covered 2015 – 2023 (9 years)
Curriculum years 1st – 5th year
Language French
Vocabulary size 29,284 words
Questions with clinical context 914 (13.5%)
Avg. context length 102.7 words (range: 11–1,049)
5-option questions (A–E) 6,198 (91.5%)
4-option questions (A–D) 573 (8.5%)
Single-answer questions (MCQU) 1,465 (21.6%)
Multi-answer questions (MCQM) 5,306 (78.4%)
— of which multi-answer 5,273 (77.9%)
— of which no correct answer 33 (0.5%)
Average correct options 2.58
Understanding questions 5,323 (78.6%)
Reasoning questions 1,448 (21.4%)
Avg. question length 11.9 words
Avg. answer option length 6.6 words
Normale sessions 3,891 (57.5%)
Rattrapage sessions 2,860 (42.2%)
Exceptionnelle sessions 20 (0.3%)

Distribution by Question Type

Split Total MCQU % MCQU % of all MCQU MCQM % MCQM % of all MCQM
Full dataset 6,771 1,465 21.6% 100.0% 5,306 78.4% 100.0%
Train 4,739 1,029 21.7% 70.2% 3,710 78.3% 69.9%
Validation 677 152 22.5% 10.4% 525 77.5% 9.9%
Test 1,355 284 21.0% 19.4% 1,071 79.0% 20.2%

Note: MCQU = single-answer questions (1 correct option). MCQM = multiple-answer questions (2 or more correct options) and questions with no correct answer (n=33, 0.5%). The 33 no-answer questions are labeled MCQM and placed exclusively in the test split to stress-test model behavior when no correct option exists.

Distribution by Cognitive Focus

Each question is additionally annotated with a cognitive focus label indicating whether it primarily requires Understanding (recall of medical knowledge) or Reasoning (multi-step clinical reasoning or integration of information). Labels were produced automatically using GPT-4o (gpt-4o-2024-08-06), following the prompt strategy of Zuo et al. (MedXpertQA) and MediQAl.

Split Total Understanding % Reasoning %
Full dataset 6,771 5,323 78.6% 1,448 21.4%
Train 4,739 3,738 78.9% 1,001 21.1%
Validation 677 509 75.2% 168 24.8%
Test 1,355 1,076 79.4% 279 20.6%

Note: Reasoning questions are correlated with the presence of a clinical context, but the label is not reducible to it: 45.4% of Reasoning questions (658) have no clinical context, indicating the annotation captures reasoning demand beyond context presence alone.

Distribution by Curriculum Year

Curriculum Year Total % Normale Rattrapage Exceptionnelle Train Nor.(Tr) Ratt.(Tr) Exc.(Tr) Val Nor.(V) Ratt.(V) Exc.(V) Test Nor.(Te) Ratt.(Te) Exc.(Te)
1st Year 795 11.7% 492 303 0 556 344 212 0 80 44 36 0 159 104 55 0
2nd Year 484 7.1% 237 247 0 339 176 163 0 48 20 28 0 97 41 56 0
3rd Year 548 8.1% 343 205 0 385 236 149 0 54 36 18 0 109 71 38 0
4th Year 1,842 27.2% 1,057 785 0 1,289 721 568 0 185 114 71 0 368 222 146 0
5th Year 3,102 45.8% 1,762 1,320 20 2,170 1,240 918 12 310 176 132 2 622 346 270 6
Total 6,771 100% 3,891 2,860 20 4,739 2,717 2,010 12 677 390 285 2 1,355 784 565 6

Distribution by Medical Specialty

Specialty Total % of Total Train Val Test
Synthèse thérapeutique et raisonnement clinique 577 8.52% 403 58 116
Biochimie clinique 484 7.15% 339 48 97
Gynécologie Obstétrique 479 7.08% 335 48 96
Maladies de l'enfant 447 6.60% 312 45 90
Ophtalmologie 399 5.89% 279 40 80
Santé mentale 397 5.86% 278 39 80
Maladies du système nerveux 386 5.70% 270 39 77
Immunopathologie 366 5.41% 256 37 73
Médecine Sociale et Santé Publique 311 4.59% 218 31 62
Maladies de l'appareil digestive 302 4.46% 212 30 60
ORL 270 3.99% 189 27 54
Dermatologie 261 3.85% 183 26 52
Histologie/Embryologie 248 3.66% 173 25 50
Médecine Légale et Déontologie 240 3.54% 168 24 48
Biophysique 239 3.53% 167 24 48
APPAREIL CARDIO-VASCULAIRE 234 3.46% 164 23 47
Psycho-sociologie 231 3.41% 162 23 46
Urgences et réanimation 207 3.06% 145 21 41
Endocrinologie - Diabétologie 140 2.07% 98 14 28
Néphrologie 139 2.05% 97 14 28
Traumatologie-Orthopédie 119 1.76% 83 12 24
Rhumatologie 70 1.03% 49 7 14
Urologie 60 0.89% 42 6 12
Biologie 50 0.74% 35 5 10
Hématologie Clinique 34 0.50% 24 3 7
Ethique 25 0.37% 18 2 5
Histoire de la médecine 17 0.25% 12 2 3
Maladies de l'appareil respiratoire 12 0.18% 9 1 2
Stage d'Immersion 10 0.15% 7 1 2
Radiothérapie 10 0.15% 7 1 2
Anatomie Pathologique II et III 7 0.10% 5 1 1
Total 6,771 100% 4,739 677 1,355

Distribution by Exam Year

Year Total Normale Rattrapage Exceptionnelle Train Nor.(Tr) Ratt.(Tr) Exc.(Tr) Val Nor.(V) Ratt.(V) Exc.(V) Test Nor.(Te) Ratt.(Te) Exc.(Te)
2015 323 183 140 0 215 121 94 0 34 17 17 0 74 45 29 0
2016 916 498 418 0 651 346 305 0 88 55 33 0 177 97 80 0
2017 1,133 596 537 0 803 423 380 0 109 51 58 0 221 122 99 0
2018 1,180 751 429 0 801 510 291 0 117 77 40 0 262 164 98 0
2019 278 195 63 20 193 139 42 12 34 22 10 2 51 34 11 6
2020 205 167 38 0 148 122 26 0 17 12 5 0 40 33 7 0
2021 974 478 496 0 688 328 360 0 85 49 36 0 201 101 100 0
2022 1,150 636 514 0 800 454 346 0 140 75 65 0 210 107 103 0
2023 612 387 225 0 440 274 166 0 53 32 21 0 119 81 38 0
Total 6,771 3,891 2,860 20 4,739 2,717 2,010 12 677 390 285 2 1,355 784 565 6

Dataset Structure

Files

File Format Records Description
MoroccanMedMCQA-FR.json JSON 6,771 Full dataset
MoroccanMedMCQA-FR.csv CSV 6,771 Full dataset
MoroccanMedMCQA-FR_train.json/csv JSON/CSV 4,739 (70%) Train split
MoroccanMedMCQA-FR_val.json/csv JSON/CSV 677 (10%) Validation split
MoroccanMedMCQA-FR_test_no_answers.json/csv JSON/CSV 1,355 (20%) Test split (answers withheld)

Splits

The dataset is split using stratified sampling by medical specialty, resulting in proportional representation of specialties, curriculum years, and exam years across the train, validation, and test splits:

Split Records Percentage
Train 4,739 70%
Validation 677 10%
Test 1,355 20%

Note: The test split is provided without correct answers to preserve benchmark integrity. Researchers wishing to evaluate on the test set should submit their predictions to hamza.aouadi@usmba.ac.ma.

Data Fields

Example 1 — Question with clinical context and 5 options:

{
    "id": 26,
    "question_type": "MCQM",
    "question_focus": "Understanding",
    "curriculum_year": "1er année",
    "exam_session": {
        "fmp": "Fez",
        "year": 2015,
        "session_type": "Rattrapage"
    },
    "speciality": "Biologie",
    "context": "La culture des cellules tumorales a été réalisée dans un milieu nutritif et riche en CO2. Après quelques jours de culture, les cellules reçoivent le gène P53 et sont séparées en 2 groupes. Le groupe 1 reçoit par la suite du ca2+ et le produit Arf (il bloque la formation du complexe MDM2-P53) et on constate que les cellules de ce groupe ne dépassent pas la phase G2. Le groupe 2 reçoit les facteurs de croissance et l'anticorps antiP53 et on constate que les cellules de ce groupe prolifèrent.",
    "question": "La P53",
    "answers": {
        "a": "Agit négativement sur la croissance cellulaire, lorsqu'elle est libre",
        "b": "Agit sur les cellules par ubiquitinilation des petites protéines P21 ,et P27 (inhibitrices de la phase G1)",
        "c": "Bloque le cycle cellulaire par action sur le complexe cycline B-CDK1 (activateur de la phase G2)",
        "d": "Le MDM2 est inhibiteur direct de la P53",
        "e": "Le produit Arf entraîne un endommagement de l'ADN parce qu'il bloque le complexe MDM2-P53"
    },
    "correct_answers": ["a", "c", "d"]
}

Example 2 — Question without context and without option E:

{
    "id": 1,
    "question_type": "MCQM",
    "question_focus": "Understanding",
    "curriculum_year": "1er année",
    "exam_session": {
        "fmp": "Fez",
        "year": 2015,
        "session_type": "Normale"
    },
    "speciality": "Biologie",
    "context": null,
    "question": "Les connexines sont des protéines de communication des cellules de l'oreille interne. Chaque jonction est composée de connexons qui sont constitués de connexines formant ainsi un canal, dont la nécessité est capitale dans de nombreux processus physiologiques",
    "answers": {
        "a": "Les connexines sont des molécules de jonction serrée (tight)",
        "b": "Les connexines sont des molécules de jonction communicante (gap)",
        "c": "Les connexines sont des molécules de communication intercellulaire paracrine",
        "d": "Les connexines permettent une communication intracytoplasmique des cellules adjacentes",
        "e": null
    },
    "correct_answers": ["b", "d"]
}
Field Type Description
id integer Unique question identifier
question_type string Question type: MCQU (single correct answer) or MCQM (multiple correct answers, or no correct answer)
question_focus string Cognitive focus: Understanding (knowledge recall) or Reasoning (multi-step clinical reasoning). Automatically annotated with GPT-4o.
curriculum_year string Medical curriculum year (1st–5th)
exam_session.fmp string Faculty of Medicine location ("Fez")
exam_session.year integer Exam year (2015–2023)
exam_session.session_type string Session type (Normale / Rattrapage / Exceptionnelle)
speciality string Medical specialty
context string or null Clinical context/vignette (null if absent)
question string Question text
answers.a–e string or null Answer options (null if option not present)
correct_answers list List of correct option letters in lowercase ([] if none)

Usage

Repository File Structure 

hamzaaouadi/MoroccanMedMCQA-FR/
├── data/
│   ├── full/
│   │   ├── MoroccanMedMCQA-FR.json
│   │   └── MoroccanMedMCQA-FR.csv
│   ├── train/
│   │   ├── MoroccanMedMCQA-FR_train.json
│   │   └── MoroccanMedMCQA-FR_train.csv
│   ├── val/
│   │   ├── MoroccanMedMCQA-FR_val.json
│   │   └── MoroccanMedMCQA-FR_val.csv
│   └── test/
│       ├── MoroccanMedMCQA-FR_test_no_answers.json
│       └── MoroccanMedMCQA-FR_test_no_answers.csv
└── README.md

Loading with 🤗 Datasets 

from datasets import load_dataset

# Load full dataset
dataset = load_dataset("hamzaaouadi/MoroccanMedMCQA-FR",
                       data_files="data/full/MoroccanMedMCQA-FR.json")

# Load train split
train = load_dataset("hamzaaouadi/MoroccanMedMCQA-FR",
                     data_files="data/train/MoroccanMedMCQA-FR_train.json")

# Load validation split
val = load_dataset("hamzaaouadi/MoroccanMedMCQA-FR",
                   data_files="data/val/MoroccanMedMCQA-FR_val.json")

# Load test split (no answers)
test = load_dataset("hamzaaouadi/MoroccanMedMCQA-FR",
                    data_files="data/test/MoroccanMedMCQA-FR_test_no_answers.json")

Loading with pandas (CSV) 

import pandas as pd

# Load full dataset
df = pd.read_csv("data/full/MoroccanMedMCQA-FR.csv", sep=";", encoding="utf-8-sig")

# Load train split
train_df = pd.read_csv("data/train/MoroccanMedMCQA-FR_train.csv", sep=";", encoding="utf-8-sig")

# Filter by specialty
ophtalmologie = df[df["speciality"] == "Ophtalmologie"]

# Filter by curriculum year
year5 = df[df["curriculum_year"] == "5eme année"]

# Filter by session type
normale = df[df["session_type"] == "Normale"]

# Filter by question type
mcqm = df[df["question_type"] == "MCQM"]
mcqu = df[df["question_type"] == "MCQU"]

# Filter by cognitive focus
reasoning = df[df["question_focus"] == "Reasoning"]
understanding = df[df["question_focus"] == "Understanding"]

Loading JSON directly 

import json

with open("data/full/MoroccanMedMCQA-FR.json", "r", encoding="utf-8") as f:
    data = json.load(f)

# Example: get all multi-answer questions (MCQM)
mcqm = [q for q in data if q["question_type"] == "MCQM"]
print(f"MCQM questions: {len(mcqm)}")

# Example: get all single-answer questions (MCQU)
mcqu = [q for q in data if q["question_type"] == "MCQU"]
print(f"MCQU questions: {len(mcqu)}")

# Example: get all reasoning questions
reasoning = [q for q in data if q["question_focus"] == "Reasoning"]
print(f"Reasoning questions: {len(reasoning)}")

# Example: get questions with clinical context
with_context = [q for q in data if q["context"] is not None]
print(f"Questions with context: {len(with_context)}")

Data Collection and Quality

Source

All questions were collected from official past examinations of the Faculty of Medicine and Pharmacy of Fès (FMPF), Morocco, with official corrections provided by the faculty. No crowdsourcing or AI generation was used.

Quality Control

  • Duplicate questions removed (22 duplicates identified and removed)
  • Answer options normalized and validated
  • Correct answers verified against official faculty corrections
  • Session types and specialty labels standardized

Limitations

  • Dataset covers only one Moroccan faculty (FMPF, Fès) — results may not generalize to other Moroccan or Francophone medical curricula
  • Exam years 2015–2023 only — older content not included
  • 33 questions (0.5%) have no correct answer among the provided options; these are labeled MCQM and placed exclusively in the test split to evaluate whether models can handle the absence of a correct option.
  • 7 medical specialties have fewer than 50 questions, which may affect the reliability of per-specialty evaluation results

Access and License

License

This dataset is released under CC BY-NC 4.0. It is available exclusively for non-commercial academic research.

Gated Access

Access to MoroccanMedMCQA-FR is gated. To request access:

  1. Click the "Access repository" button above
  2. Fill in your name, institutional affiliation, and intended use
  3. Agree to the Data Usage Agreement
  4. Requests are reviewed manually within 14 business days

For questions or issues, contact: hamza.aouadi@usmba.ac.ma

Data Usage Agreement

By accessing this dataset, you agree to:

  • Use the dataset solely for non-commercial academic research
  • Not redistribute, share, or sublicense the dataset
  • Cite the original paper in any resulting publication
  • Not use the dataset to train or evaluate commercial AI systems

Citation

If you use MoroccanMedMCQA-FR in your research, please cite:

@dataset{aouadi2025moroccanmedmcqa-fr,
  author    = {AOUADI, Hamza and NFAOUI, El Habib and ELGAROUANI, Said},
  title     = {MoroccanMedMCQA-FR: A French-Language Moroccan Medical Multiple-Choice QA Benchmark},
  year      = {2025},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/hamzaaouadi/MoroccanMedMCQA-FR},
  note      = {Paper: [PLACEHOLDER — to be updated upon publication]}
}

The BibTeX entry will be updated with the full paper reference upon publication.

Acknowledgements

This dataset was compiled from official past examinations of the Faculty of Medicine and Pharmacy of Fès (FMPF), Sidi Mohammed Ben Abdellah University (USMBA), Morocco. We thank the faculty for making these resources available for academic research.
Downloads last month
28
Total file size:
18.8 MB