sinhala-nlp/sinhala-1.5B-corpus

Sinhala-1.5B-corpus is a large-scale monolingual corpus for Sinhala, released as part of the ACL 2025 paper "Sinhala Encoder-only Language Models and Evaluation".

The corpus was created to address the limited availability of large-scale pre-training resources for Sinhala. It combines Sinhala text from a range of domains, including web content, news, social media, books, educational material and government documents.

The corpus was used to train the Sinhala encoder-only language models introduced in our paper: Raja (Sinhala-BERT), Koliya (Sinhala-RoBERTa) and Mahasen (Sinhala-ELECTRA).

Dataset Description

Sinhala remains comparatively under-resourced in NLP despite being spoken by more than 17 million people. Existing multilingual pre-training corpora contain relatively small amounts of Sinhala data, limiting the quality of Sinhala representations learned by multilingual language models.

To provide a larger resource for Sinhala language modelling, we combined six publicly available datasets:

Source	Tokens	Documents	Original Licence
HPLT 2.0	934,236,876	1,152,703	CC0
FineWeb2	434,560,077	1,077,501	ODC-By 1.0
NSina	94,394,362	486,932	CC BY-NC-SA 4.0
FacebookDecadeCorpora (FDC)	5,402,768	364,402	CC BY 4.0
SinMin	104,428,504	313,910	CC BY
SemiSOLD	1,938,756	107,210	CC BY 4.0

In total, the source datasets contain more than 1.5 billion tokens and more than 3.5 million documents. A token in these statistics is defined as any continuous sequence of non-whitespace characters.

Source Data

HPLT 2.0
A multilingual corpus derived from the Internet Archive and Common Crawl, covering 75 languages including Sinhala.

FineWeb2
A large multilingual web corpus constructed from Common Crawl snapshots between 2013 and 2024.

NSina
A Sinhala news corpus collected from ten news websites in Sri Lanka. The corpus includes material from a range of news sources.

FacebookDecadeCorpora (FDC)
A social-media corpus collected from Sri Lankan Facebook pages between 2010 and 2020, covering areas including politics, media and celebrities.

SinMin
A Sinhala corpus containing modern and historical texts from multiple genres. Sources include newspapers, magazines, school textbooks, the Mahawansa, Sinhala Wikipedia, Sri Lankan government gazettes and subtitles.

SemiSOLD
A collection derived from Sinhala Twitter data originally collected for offensive-language research. Only instances labelled as non-offensive were included in this corpus.

Dataset Structure

The Hugging Face version contains the following fields:

text    : string
source  : string

• text contains the textual content.
• source identifies the source dataset from which the instance originated.

The dataset is provided as a training corpus without predefined validation or test splits.

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("sinhala-nlp/Sinhala-1.5B-corpus")

print(dataset)
print(dataset["train"][0])

You can access the text directly using:

texts = dataset["train"]["text"]

or filter instances according to their original source:

hplt = dataset["train"].filter(
    lambda example: example["source"] == "hplt_2"
)

Intended Uses

The corpus is primarily intended for research involving Sinhala natural language processing, including:

• pre-training Sinhala language models;
• continued pre-training or language adaptation of multilingual models;
• representation learning;
• language modelling;
• developing tokenisers for Sinhala;
• studying low-resource language modelling;
• constructing Sinhala NLP systems and resources.

The corpus was originally used to train encoder-only transformer language models, but it may also be useful for other language-modelling approaches.

Languages

The primary language of the corpus is Sinhala (si).

Since several components originate from large-scale web crawls and other naturally occurring sources, individual documents may contain English, other languages, code-mixed text, markup or other noisy content.

Data Quality

The corpus combines data from several independently collected resources and therefore varies in style, domain and quality.

The source material includes web-crawled text and social-media content. Consequently, users should expect some degree of:

• duplicated or near-duplicated content;
• non-Sinhala text and code-mixed material;
• web artefacts and formatting noise;
• spelling and grammatical variation;
• outdated information;
• potentially offensive or otherwise undesirable content;
• factual inaccuracies present in the original sources.

Although only non-offensive instances from SemiSOLD were included, this should not be interpreted as a guarantee that the complete corpus is free from offensive, harmful or sensitive material.

Biases and Limitations

The corpus should not be considered a representative sample of all Sinhala language use.

Its composition is influenced by the availability of publicly accessible digital Sinhala resources. Web pages and news articles constitute a substantial proportion of the corpus, while spoken language, conversational language and some specialised domains are comparatively under-represented.

The original study did not conduct a systematic analysis of potential social, demographic or other biases within the corpus. Models trained on the dataset may therefore reproduce biases, stereotypes, factual errors or other undesirable patterns contained in the underlying data.

Users deploying models trained using this corpus should conduct appropriate evaluation for their intended application, particularly for high-impact or sensitive settings.

Licensing

This corpus combines material released under multiple licences. The original licence of each component dataset is retained:

• HPLT 2.0: CC0
• FineWeb2: ODC-By 1.0
• NSina: CC BY-NC-SA 4.0
• FacebookDecadeCorpora: CC BY 4.0
• SinMin: CC BY
• SemiSOLD: CC BY 4.0

Users are responsible for ensuring that their use of individual components complies with the corresponding source licence. In particular, some components impose attribution, share-alike and/or non-commercial requirements.

Please consult the original datasets and their licence documentation before redistribution or commercial use.

Citation

If you use this corpus in your research, please cite:

@inproceedings{ranasinghe-etal-2025-sinhala,
    title = "{S}inhala Encoder-only Language Models and Evaluation",
    author = "Ranasinghe, Tharindu  and
      Hettiarachchi, Hansi  and
      Pathirana, Nadeesha  and
      Premasiri, Damith  and
      Uyangodage, Lasitha  and
      Arachchige, Isuri  and
      Plum, Alistair  and
      Rayson, Paul  and
      Mitkov, Ruslan",
    booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2025",
    address = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.acl-long.422/",
    doi = "10.18653/v1/2025.acl-long.422",
    pages = "8623--8636",
    ISBN = "979-8-89176-251-0"
}

Related Resources

The corpus was introduced together with three Sinhala encoder-only language models and the SINHALA-GLUE benchmark for Sinhala natural language understanding.

Paper: Sinhala Encoder-only Language Models and Evaluation, ACL 2025.

The associated work also releases the pre-trained models, fine-tuned models, Sinhala-GLUE datasets and evaluation resources through the Sinhala NLP organisation.