---
license: cc-by-nc-4.0
language:
- es
- eu
task_categories:
- question-answering
pretty_name: SNS1064
size_categories:
- 1K<n<10K
---

# SNS1064

SNS1064 is an open-answer medical question-answering dataset built from Spanish
medical guidebooks published by the Sistema Nacional de Salud (SNS). It transforms
unstructured clinical guideline text into 1,064 structured question-answer-evidence
instances, each separating the question asked, a short answer, and the evidence
passage from the guideline that supports it.

The dataset is provided in two languages:

- **`es/`**: the original Spanish instances.
- **`eu/`**: a Basque translation of the same instances (machine-translated from
  the Spanish originals, with automatic checks for truncation and degenerate
  repetition applied to the translation pipeline).

Both language versions share the same instance ids, schema, and train/dev/test
split, so a given `id` refers to the same underlying question in either language.

## Dataset structure

Each split is a JSON Lines file (`train.jsonl`, `dev.jsonl`, `test.jsonl`) under
the corresponding language directory (`es/` or `eu/`). Each line is one instance:

```json
{
  "id": "sns1064_00000",
  "source": "SNS1064",
  "guidebook": "ansiedad.txt",
  "topic": "...",
  "question": "...",
  "short_answer": "...",
  "evidence": "...",
  "split": "train"
}
```

- `id`: a globally unique instance identifier, shared across the `es` and `eu` versions.
- `source`: dataset name (always `"SNS1064"`).
- `guidebook`: the source clinical guideline file the instance was extracted from.
- `topic`: the clinical question the guideline recommendation addresses.
- `question`: the specific sub-question asked (e.g. certainty of evidence, cost-effectiveness).
- `short_answer`: a short, direct answer.
- `evidence`: the guideline passage that supports the answer.
- `split`: one of `train`, `dev`, `test`.

## Splits

| Split | Instances |
|---|---|
| train | 876 |
| dev | 63 |
| test | 125 |
| **Total** | **1,064** |

## Intended use

SNS1064 is designed to support retrieval-augmented generation (RAG) and
open-answer question-answering research in the medical domain, particularly for
Spanish and Basque, a low-resource language pair. It was built as part of a
Master's thesis evaluating RAG pipelines for medical QA in Spanish and Basque.

## License

This dataset is released under the [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/) license.

## Citation

If you use this dataset, please cite the accompanying thesis (citation details to
be added upon publication).