JMasr/balidea-medquad-qa-gl

---
language:
- en
- gl
license: apache-2.0
task_categories:
- text-classification
- question-answering
tags:
- medical
- translation
- galician
- medquad
- synthetic
size_categories:
- 10K<n<100K
---

# balidea-medquad-qa-gl

Galician translation of the [MedQuAD](https://github.com/abachaa/MedQuAD) medical question-answering dataset. Contains 16,407 question-answer pairs translated from English to Galician using an automatic translation pipeline.

## Dataset splits

| Split | Rows |
|-------|-----:|
| train | 13,125 |
| validation | 1,641 |
| test | 1,641 |
| **Total** | **16,407** |

## Features

- `text` — Medical question or answer in Galician (`string`)
- `labels` — Classification label (`int64`)

## Translation pipeline

Translations were produced by a dual-engine pipeline designed to balance speed, fluency, and domain accuracy:

1. **Dual-engine translation** — Each sentence is translated in parallel by two models:
- [Helsinki-NLP OPUS-MT](https://huggingface.co/Helsinki-NLP) — a lightweight, fast neural MT model
- TranslateGemma 12B — a large language model with stronger contextual understanding

2. **Arbitration** — A second TranslateGemma 12B instance acts as an arbitrator, comparing both outputs and synthesizing a final translation that prioritises natural fluency and adherence to [ILG-RAG](https://ilg.usc.gal/) orthographic standards for Galician.

3. **Quality scoring** — Every translation is scored with [COMETKiwi](https://huggingface.co/Unbabel/wmt22-cometkiwi-da) (reference-free quality estimation). Sentences below the quality threshold are flagged for manual review.

## Source dataset

Based on [MedQuAD](https://github.com/abachaa/MedQuAD) (Medical Question Answering Dataset) by Ben Abacha and Demner-Fushman (2019), covering 37 medical categories sourced from NIH websites.