tags:
- low-resource-language
- mixed-script
pretty_name: Sinhala Mixed-Script Benchmark Dataset
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: data
path: data/data*
dataset_info:
features:
- name: text
dtype: string
splits:
- name: data
num_examples: 500
language:
- si
Dataset Description
This dataset contains 500 authentic Sinhala mixed-script sentences in which Unicode Sinhala and Romanized Sinhala scripts naturally co-occur within the same text. The sentences were collected from social media platforms, blogs and online forums, preserving real-world script-switching patterns without synthetic manipulation. The dataset is designed to support research on language model robustness, script mixing, and low-resource multilingual NLP, particularly under mixed-script conditions.
This dataset was introduced and used in the following benchmark study:
Rajapakse, M., & Weerasinghe, R. (2025). Script Sensitivity: Benchmarking Language Models on Unicode, Romanized and Mixed-Script Sinhala. arXiv:2601.14958v2. https://arxiv.org/abs/2601.14958v2
If you use this dataset, please cite the above paper.
Citation
@article{rajapakse2026comprehensive,
title={A Comprehensive Benchmark of Language Models on Unicode and Romanized Sinhala},
author={Rajapakse, Minuri and Weerasinghe, Ruvan},
journal={arXiv preprint arXiv:2601.14958},
year={2026}
}
Dataset Structure
The dataset contains a single text field:
text: A naturally occurring Sinhala mixed-script sentence containing both Unicode and Romanized scripts.
Statistics
- Number of sentences: 500
- Mean sentence length: 9.49 ± 4.03 tokens
- Vocabulary size: 2,262 unique tokens
How to Use the Dataset
You can load the dataset using the Hugging Face datasets library:
from datasets import load_dataset
dataset = load_dataset("Minuri/sinhala-mixedscript")
example = dataset["data"][0]
print(example["text"])