MedQA-Darija-MCQ-500 — Moroccan Darija Medical MCQ Benchmark

A 500-question multiple-choice benchmark for medical reasoning in Moroccan Arabic Darija. To our knowledge this is the first publicly released medical MCQ benchmark in Darija — built specifically because no equivalent existed and Darija-medical LLM evaluation had nowhere to land.

Stat	Value
Items	500
Language	Moroccan Arabic Darija
Format	4-option MCQ (A/B/C/D, single correct)
Source stems	BrainHealthAI/MedQADataDarijaSaad test split (real Darija clinical Q&A)
Distractor generator	groq/llama-3.3-70b-versatile (free tier)
Build acceptance rate	~93% (500 accepted out of 537 attempts; 37 rejected by validation)

Schema

Field	Type	Description
question	string	Patient/clinical question in Darija (includes original context when present)
choices	list[string]	Exactly 4 options, order randomised
correct_letter	string	A / B / C / D
specialty	string	Source-reported medical specialty (Arabic)
urgency	string	Faible / Moyen / Élevé / unknown (source labels)
source_pair_id	string	Reference id back to the Saad source row

How it was built

1. Load BrainHealthAI/MedQADataDarijaSaad test split (813 real Darija Q&A pairs).
2. For each item, ask groq/llama-3.3-70b-versatile to generate 3 plausible-but-wrong distractors in Darija, of comparable length, using a strict JSON-only schema. The correct answer is the original Saad answer.
3. Validate each distractor (deterministic gates, no LLM judge):
   • non-empty, ≥ 8 chars, ≤ 600 chars
   • length ratio with the correct answer in [0.4, 2.5]
   • not a substring of the correct answer (or vice versa) — catches near-duplicates
   • distinct from the other distractors of the same item
4. Randomise the position of the correct answer (uniform), record correct_letter.
5. Items failing any gate are dropped (37 of 537 attempts).

Suggested evaluation prompt

For a fair comparison across models:

System: You are a medical exam assistant. Choose the single best option.
Reply with ONLY the letter (A, B, C, or D), no explanation.
User: {question}
A. {choices[0]}
B. {choices[1]}
C. {choices[2]}
D. {choices[3]}

Parse the model's first letter token, compare to correct_letter. Compute accuracy + calibration (ECE if you have access to logits).

Intended use

• Evaluate medical LLMs on Moroccan Darija reasoning
• Stress-test multilingual medical models (most are English-only on benchmarks)
• Compare your model against meta-llama/Llama-3.1-8B-Instruct, BioMistral/BioMistral-7B, epfl-llm/meditron-7b, etc.

Honest limitations (please read)

• Distractors are LLM-generated, not physician-validated. Some distractors may be closer in meaning to the correct answer than intended, making certain items effectively ambiguous. A physician-review pass on a 100-item sample is planned and will be released as a *-curated config when done.
• Darija orthography varies. Moroccan Darija has no single standard; spelling and loanword conventions differ from author to author. The benchmark inherits the Saad source's conventions.
• Specialty distribution mirrors the source. General medicine and gynaecology are over-represented (this is the structure of the Saad source). Use specialty-stratified splits if you need uniform coverage.
• Built atop a single source (Saad). A fully independent Darija benchmark would require fresh medical Q&A authored from scratch by clinicians — that's a separate research effort.
• Free-tier LLM generation. The distractors reflect any biases of llama-3.3-70b-versatile. Models from the same family may have a small but real advantage here (the generator and the evaluatee share priors).

References / Related work

The benchmark format and the surrounding evaluation methodology draw on:

• MedQA (USMLE) (4-option clinical MCQ benchmark, format inspiration) — Jin, D. et al. (2021). What Disease does this Patient Have? A Large-scale Open Domain Question Answering Dataset from Medical Exams. Applied Sciences. arXiv:2009.13081
• MedMCQA (large-scale multi-subject medical MCQ) — Pal, A. et al. (2022). MedMCQA: A Large-scale Multi-Subject Multi-Choice Dataset for Medical domain Question Answering. CHIL 2022. arXiv:2203.14371
• PubMedQA (biomedical question answering) — Jin, Q. et al. (2019). PubMedQA: A Dataset for Biomedical Research Question Answering. EMNLP 2019. arXiv:1909.06146
• MMLU (incl. clinical_knowledge / medical_genetics / professional_medicine / college_medicine subsets) — Hendrycks, D. et al. (2021). Measuring Massive Multitask Language Understanding. ICLR 2021. arXiv:2009.03300
• Llama 3 (the distractor-generation model family) — Grattafiori, A. et al. (2024). The Llama 3 Herd of Models. arXiv:2407.21783
• AraBERT (Arabic NLP foundation; relevant baseline for Arabic-family medical NLP) — Antoun, W. et al. (2020). AraBERT: Transformer-based Model for Arabic Language Understanding. OSACT 2020. arXiv:2003.00104
• MEDITRON-7B/70B (open medical LLM for English comparison) — Chen, Z. et al. (2023). arXiv:2311.16079
• BioMistral-7B (open medical LLM for English/multilingual comparison) — Labrak, Y. et al. (2024). arXiv:2402.10373

Source data: BrainHealthAI/MedQADataDarijaSaad (the held-out test split of which became the question stems for this benchmark).

License

CC BY-NC 4.0 — research / non-commercial.

Citation

@misc{medqa_darija_mcq_500_2026,
  author       = {BrainHealthAI},
  title        = {{MedQA-Darija-MCQ-500: A Moroccan Darija Medical MCQ Benchmark}},
  year         = 2026,
  publisher    = {HuggingFace},
  howpublished = {\url{https://huggingface.co/datasets/BrainHealthAI/MedQA-Darija-MCQ-500}},
}

Disclaimer

Not for clinical decision-making. Research benchmark only.