SAMSUM-eu dataset for Summarization in Basque.

SAMSUM-eu was created by automatically translating SAMSum, a human-annotated dialogue dataset for abstractive summarization, using a proprietary document-level MT system based on Llama-eus-8B. We then filtered out examples with incomplete translations or non-Basque outputs. The translated test set was further refined by a native speaker to obtain 100 high-quality, manually curated test examples. In total, we obtained 11,313 training examples, 636 validation examples, and 100 manually curated test examples for evaluation.

📝 Paper: Sub-1B Language Models for Low-Resource Languages: Training Strategies and Insights for Basque accepted in 5TH MULTILINGUAL REPRESENTATION LEARNING (MRL) WORKSHOP 2025 (EMNLP)

Acknowledgments

The creation of this dataset has been partially funded by the Basque Government (ICL4LANG project, grant no. KK-2023/00094) and the European Union (EFA 104/01-LINGUATEC IA project, INTERREG POCTEFA 2021-2027 program). Finally, we thank Idoia Davila Uzkudun for her contributions to manual data curation and evaluation.

Citation

If you use this dataset please cite the following paper:

@inproceedings{urbizu2025sub,
  title={Sub-1B Language Models for Low-Resource Languages: Training Strategies and Insights for {B}asque},
  author={Urbizu, Gorka and Corral, Ander and Saralegi, Xabier and San Vicente, I{\~n}aki},
  booktitle={Proceedings of the 5th Workshop on Multilingual Representation Learning (MRL 2025)},
  pages={519--530},
  year={2025}
}

Contact

• Gorka Urbizu ([email protected])
• Xabier Saralegi ([email protected])