janani-rane/SiQuAD

Sinhala SQuAD Dataset

This dataset is a translation of the SQuAD v1.0 dataset into Sinhala using the Google Translate API. It consists of 16,000 question-answer pairs, with 13,000 training pairs and 1,250 test/dev pairs. The dataset is cleaned and validated to ensure the quality of the translations.

Dataset Details

• Size: 16,000 QA pairs
  • Train: 13,000 pairs
  • Test/Dev: 1,250 pairs
• Language: Sinhala
• Source: The dataset was derived from the original SQuAD v1.0 dataset.
• Preprocessing: The data was translated using the Google Translate API, followed by cleaning and validation of the translations.

Performance

The best performance achieved on this dataset was obtained by fine-tuning the XLM-R Large model on the Sinhala-translated data. The model achieved the following performance metrics:

• F1 Score: 73.29 on the test set

Usage

You can use this dataset for training and evaluating extractive question answering models in Sinhala. It can be easily loaded using the Hugging Face datasets library.

from datasets import load_dataset

dataset = load_dataset("janani-rane/SiQuAD")

Citation

If you use this dataset, please cite the following paper:

Ranasinghe, J., & Weerasinghe, R. (2025). Question Answering in a Low-Resource Language: Dataset and Deep Learning Adaptations for Sinhala. In: Hadjali, A., Maiorana, E., Gusikhin, O., & Sansone, C. (eds) Deep Learning Theory and Applications. DeLTA 2025. Communications in Computer and Information Science, vol 2627. Springer, Cham. https://doi.org/10.1007/978-3-032-04339-9_22

BibTeX

@inproceedings{ranasinghe2025sinhalaqa,
  author    = {Janani Ranasinghe and Ruvan Weerasinghe},
  title     = {Question Answering in a Low-Resource Language: Dataset and Deep Learning Adaptations for Sinhala},
  booktitle = {Deep Learning Theory and Applications (DeLTA 2025)},
  series    = {Communications in Computer and Information Science},
  volume    = {2627},
  publisher = {Springer},
  year      = {2025},
  doi       = {10.1007/978-3-032-04339-9_22}
}