MedQA-Darija-MultiLingual

The largest open trilingual medical Q&A dataset with directly-playable speech audio for English, French, and Moroccan Darija.

A research dataset for the BRAIN HEALTH initiative, designed for multilingual medical NLP, low-resource speech recognition, healthcare chatbots, and clinical education tools targeting Morocco and the broader Maghreb region.

Dataset is currently in scientific validation phase. After programmatic validation (Stage 1 LOF outlier detection, Stage 2 Causal-DAG consistency, Stage 3 linguistic & medical NER coverage), it will be reviewed by licensed medical professionals before peer-reviewed publication.

At a glance

Configurations

default — Synthetic (104,139 rows)

Trilingual synthetic Q&A generated by multiple LLM providers (Mistral, Cerebras, Groq, GPT-4o-mini, OpenRouter). Every row carries six audio cells (Q+A in EN/FR/Darija), playable directly in the Hugging Face Viewer.

real — Real-world scraped (163,488 rows)

Real medical Q&A scraped from the open medical web. Two sub-sources:

• scraped_icliniq (3,383 rows) — patient questions and licensed-doctor answers from publicly available iCliniq.com pages, manually trilingualized and audio-rendered.
• open_medical_corpus (160,105 rows) — open medical corpora translated trilingually.

iCliniq rows appear first in the real split; both sub-sources share identical schema with audio columns (audio cells are populated where speech files are available, else null).

Schema (per row)

A single row contains the same medical content in three languages plus six speech audio cells:

The Hugging Face Viewer renders the six audio columns as inline players, so a reviewer can read the trilingual text and listen to the spoken version side-by-side.

Data quality & cleaning

Politeness-formula stripping

We removed boilerplate that does not carry clinical content (research-grade preprocessing):

• Greeting prefixes: "Hi/Hello/Bonjour/أهلا [doctor]" at the start of an answer
• "Welcome to icliniq.com" salutations
• Closing wishes: "Take care", "Hope this helps", "Best regards", "Prenez soin de vous", "تمنياتي ليك بالشفاء", etc.

We keep clinical content fully intact: dosages, diagnoses, mechanism explanations, clinical reasoning, and safety advice are never modified.

Three-stage validation pipeline

Every synthetic pair is required to pass:

1. Stage 1 — Representation Learning (LOF in PubMedBERT space). Embeddings of (question_en, answer_en) are scored by Local Outlier Factor; pairs with anomaly score < −0.1 are rejected as off-topic / hallucinatory.
2. Stage 2 — Causal Consistency (DoWhy DAG). A causal model symptom_complexity → diagnosis_confidence → treatment_specificity flags pairs where rich diagnoses pair with vague treatments, or vice versa.
3. Stage 3 — Linguistic & medical NER coverage. Darija fields must be ≥ 35 % Arabic characters (rejects Arabizi); French must be ≥ 60 % ASCII; medical entity density per pair is recorded.

iCliniq trilingual translations are produced via a fine-tuned MarianMT model (ychafiqui/english-to-darija-2) for Darija and a medical-prompted LLM for French. Each iCliniq pair is translated single-pair (no batched alignment errors) and passes the same three-stage validation.

Audio

• Generated with Microsoft Edge Neural TTS (ar-MA-Jamal/Mouna, fr-FR-Henri/Denise, en-US-Guy/Jenny).
• Stored as standalone MP3 referenced from each row's audio columns; Hugging Face's Audio feature handles seamless decoding and inline playback.

Specialty coverage

71 specialties spanning general medicine, surgery, paediatrics, oncology, dermatology, mental health, dentistry, radiology, etc. Each specialty has between 100 and 8,800 pairs depending on real-world data density.

(Full per-specialty table is regenerated automatically and lives in _spec_breakdown.md of the source repo.)

Intended uses

• Multilingual medical question answering (EN, FR, Darija)
• Speech-to-text / text-to-speech systems for under-resourced Moroccan Darija
• Bilingual clinical chatbots and patient-education tools
• Medical NER and relation extraction in Arabic script
• Cross-lingual retrieval and translation evaluation

Out-of-scope uses

• This dataset is not validated as a primary clinical decision-support tool.
• Audio voices are synthetic and must not be presented as actual physician voices.
• Translations and clinical content require qualified-physician review for any production-grade clinical use.

Dataset construction

Real scraping ─┐
               ├──> Trilingual translation ──> 3-stage validation ──> Edge-TTS audio ──> HF parquet + audio
LLM synth   ───┘

• Sources: open medical corpus, iCliniq scraped Q&A (publicly accessible pages, no PII).
• Translation: ychafiqui/english-to-darija-2 (MarianMT) for Darija, Mistral medical prompt for French; single-pair calls to avoid batch-alignment errors.
• Validation: PubMedBERT embeddings + LOF, DoWhy causal DAG, NER + script-character ratio gates.
• Audio: Edge Neural TTS, six voices (patient + doctor × 3 langs), MP3, stored on HF Xet.
• Cleaning: conservative politeness stripping, never touches clinical body.

Limitations

• iCliniq translations rely on automatic MT; medical reviewers will further refine before publication.
• Some specialty buckets are small (< 1,000 pairs).
• Synthetic data, while validated, may still contain factual errors not yet caught by the 3-stage pipeline.

Citation

@dataset{medqa_darija_multilingual_2026,
  title  = {MedQA-Darija-MultiLingual: a trilingual medical Q\&A dataset with audio for English, French, and Moroccan Darija},
  author = {BRAIN HEALTH},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Williamsanderson/MedQA-Darija-MultiLingual},
  note   = {Pre-publication; under medical professional review}
}

License

CC-BY-4.0 — please cite this dataset and the iCliniq attribution clause when using the scraped_icliniq subset.

Acknowledgements

• ychafiqui/english-to-darija-2 (Hugging Face) for Darija machine translation
• Microsoft Edge Neural TTS for speech rendering
• iCliniq.com for publicly-accessible medical Q&A
• Open medical corpus contributors
• BRAIN HEALTH initiative