Train split of 20 Minuten dataset (1) filtered by using the code by Aumiller et al. (2) available at https://github.com/dennlinger/summaries/tree/main with following settings:
min_length_summary=15, min_length_reference=200, length_metric="whitespace", extractiveness="fully"
After this filtering operation, the dataset was filtered again to exclude all samples with bigram overlap fraction between summary and article >= 0,76.
Finally, 5k samples were kept in this dataset. All other samples were removed.
(1): Gonzales, A. R., Spring, N., Kew, T., Kostrzewa, M., Säuberli, A., Müller, M., & Ebling, S. (2021, November). A new dataset and efficient baselines for document-level text simplification in German. In Proceedings of the Third Workshop on New Frontiers in Summarization (pp. 152-161). (2): Aumiller, D., Fan, J., & Gertz, M. (2023). On the State of German (Abstractive) Text Summarization. arXiv preprint arXiv:2301.07095.