license: gemma
language:
- am
- en
tags:
- medical
- amharic
- question-answering
- bilingual
pretty_name: Bilingual Amharic-English Medical QA
size_categories:
- 1K<n<10K
task_categories:
- question-answering
- text-generation
Bilingual Amharic–English Medical QA Dataset
A parallel medical question-answering dataset in Amharic and English,
used to fine-tune Walelign/Gemma_3_4B
and its 12B counterpart.
⚠️ Medical content notice. This dataset contains health information intended for research and educational use only. It is not a substitute for professional medical advice.
Dataset summary
Each row contains a medical question and its answer in both English and Amharic (four columns), letting a model learn the same medical facts in two languages.
| Number of rows | ~2,300 |
| Languages | Amharic (am), English (en) |
| Domain | Medical / health |
| Format | CSV |
| Answer length | Short (1–2 sentences, typical) |
Data fields
| Column | Description |
|---|---|
English_question |
The medical question in English |
English_answer |
The answer in English |
Amharic_question |
The same question in Amharic |
Amharic_answer |
The same answer in Amharic |
Example row
| Field | Content |
|---|---|
English_question |
What happens if women show syphilis signs early? |
English_answer |
Syphilis has stages where the earlier stages can be skin lesions, while the later symptoms include complications to organs including the brain and heart. |
Amharic_question |
(parallel Amharic translation of the question) |
Amharic_answer |
(parallel Amharic translation of the answer) |
How it was used for training
Each row was expanded into two training examples (English pair + Amharic pair), so ~2,300 rows produced ~4,600 examples. A 90/10 train/validation split was applied (seed 42), and each example was formatted with the Gemma chat template.
Provenance and curation
Sources. The question–answer content was compiled from a combination of: public health websites and articles, medical textbooks and clinical guidelines, expert/clinician-written material, and content adapted from an existing medical QA dataset.
Amharic translation. The Amharic questions and answers were produced by professional translation from the English source content.
Clinical review. The medical content was reviewed by a qualified health professional.
Ownership. The dataset compilation, curation, arrangement, and translation are the joint work of Walelign Tewabe Sewunetie (PhD) and Surafel L. Tilahun (PhD).
Limitations and biases
- Small size. ~2,300 pairs cover only a limited slice of medical knowledge.
- Coverage gaps. Topics not represented will be answered poorly by models trained on it.
- Translation quality. The text was professionally translated, but some specialized terms may still be inconsistent across items.
- Not a clinical reference. Answers may be simplified, outdated, or incorrect and must not be used for real medical decisions without expert review.
Intended use
Training and evaluating bilingual (Amharic/English) medical QA models for research and educational purposes only.
Citation
@misc{sewunetie_tilahun_amharic_medqa_dataset,
title = {Bilingual Amharic-English Medical QA Dataset},
author = {Sewunetie, Walelign Tewabe and Tilahun, Surafel L.},
year = {2026}
}