ErikCalcina/synthetic-multi-med-notes-ner-dataset-v1

---
pretty_name: Multilingual Synthetic Medical Notes for NER
language:
- en
- it
- es
- fr
- de
- nl
- el
- pt
- sl
license: mit
task_categories:
- token-classification
- text-classification
size_categories:
- 1K<n<10K
---

# Multilingual Synthetic Medical Notes for NER

This dataset provides multilingual synthetic clinical notes for information extraction and NER workflows.

## Dataset file

- `train.jsonl` (JSON Lines): one example per line

## Schema

Each line in `train.jsonl` contains:

- `text`: synthetic medical note text
- `language`: language of the note
- `entities`: character-level entity annotations (`text`, `label`, `start`, `end`)
- `gliner_tokenized_text`: tokenized note text for GLiNER-style training
- `gliner_entities`: token-level entity annotations aligned to `gliner_tokenized_text`

## Label scope

Entity labels cover common clinical concepts, including:

- conditions and comorbidities
- symptoms and observations
- tests, measurements, and scores
- procedures, treatments, and rehabilitation
- drugs and drug doses
- events, dates, visits, devices, and specimens

## Example

```json
{
"text": "Patient note ...",
"language": "spanish",
"entities": [{"text": "dolor", "label": "Symptom", "start": 10, "end": 15}],
"gliner_tokenized_text": ["Patient", "note", "..."],
"gliner_entities": [[5, 5, "Symptom"]]
}
```

## How synthetic data was created

- Prompt-driven generation was used to create realistic EHR-style notes across multiple languages and regions.
- Generation parameters varied across note type, specialty, patient profile, and note length to increase diversity.
- Clinical labels were embedded in the prompting process so each sample includes structured entity annotations.
- Post-generation validation kept only examples with valid JSON structure and valid entity spans that appear in the note text.
- Tokenized GLiNER fields were produced for token-level training/evaluation workflows.

## Intended use

- Training and evaluating NER / IE systems on multilingual clinical-style text
- Prototyping multilingual medical NLP pipelines

## Limitations

- Data is synthetic and may not reflect full real-world clinical variability
- Annotation quality depends on synthetic generation and preprocessing steps

## Citation

If you use this dataset, cite this repository/workspace.