README.md

---
language:
- de
license: other
size_categories:
- n<1K
task_categories:
- text-classification
- feature-extraction
- question-answering
pretty_name: UroLlmEvalSet
dataset_info:
  features:
  - name: text
    dtype: string
  - name: label
    list:
    - name: date
      dtype: string
    - name: diagnosis
      dtype: string
  splits:
  - name: eval
    num_bytes: 70066
    num_examples: 149
  download_size: 29974
  dataset_size: 70066
configs:
- config_name: default
  data_files:
  - split: eval
    path: data/eval-*
tags:
- medical
library_name: datasets
---

# UroLlmEvalSet

## Content
This dataset contains 149 short textual descriptions in German medical language about the diagnosis and medical history of patients with prostate cancer.
The prostate cancer diagnosis is not present in all of the texts and many texts also contain other tumor diagnoses.
The texts are labeled with the three-character ICD-10 codes of the diagnoses and with the date of the initial diagnosis of the tumor.
Some diagnosis dates in the text are specified only as years, while others include both year and month.
In the labels, the dates are therefore strings of the format "YYYY" or "YYYY-MM".
If there is no clearly attributable first diagnosis date for a tumor diagnosis in a text, the corresponding date value is "0".
The dataset schema is as follows:

```{json}
{
    'text': string,               # Text about diagnosis and current medical history of a patient
    'label': [                    # Multiple labels (diagnosis/date pairs) may be present for the text.
        {                         
            'diagnosis': string   # ICD-code of tumor diagnosis
            'date': string        # Date of initial diagnosis of this tumor
        }
    ]
}
```


## Loading the data set

```{python}
import datasets
datasets.load_dataset("stefan-m-lenz/UroLlmEvalSet", split="eval")
```

## Benchmarking LLMs for tumor documentation

This dataset was created for a benchmark to evaluate LLMs for tumor documentation.
The code for running the `UroLlmEval` benchmark can be found here:

https://github.com/stefan-m-lenz/UroLlmEval


## License

The data is released under terms and conditions defined in the [data usage LICENSE](LICENSE.md).
As this dataset is primarily intended for evaluation purposes, the license restricts its distribution to prevent an undisclosed inclusion in LLM training data.


## Citation and further information

Further information about the dataset and about the benchmark can be found in the following [article](https://doi.org/10.1186/s13040-025-00463-8):

```bibtex
@article{UroLlmEval_2025,
    title = {Can open source large language models be used for tumor documentation in {Germany}?—{An} evaluation on urological doctors’ notes},
    author = {Lenz, Stefan and Ustjanzew, Arsenij and Jeray, Marco and Ressing, Meike and Panholzer, Torsten},
    year = {2025},
    month = jul,
    journal = {BioData Mining},
    volume = {18},
    number = {1},
    doi = {10.1186/s13040-025-00463-8},
    url = {https://doi.org/10.1186/s13040-025-00463-8}	
}
```

If you use this dataset in your work, please consider citing the article.

You are very welcome to share feedback or join the discussion on the [Hugging Face paper page](https://huggingface.co/papers/2501.12106).