UCCIX's Irish Textual Corpus

Dataset Summary

This monolingual Irish text dataset includes data from various sources such as CulturaX, Glot500, Irish Wikipedia, providing valuable content from Irish sites and pages.

Our primary sources include CulturaX and Glot500, both of which provide important information from multilingual websites, including a subset dedicated to Irish. Additionally, we incorporate data from the Irish segment of the ga-en bitext pair of ParaCrawl v7, text sourced from Irish Wikipedia, and data from Corpora Irish. To ensure data integrity, we conduct thorough cleaning and filtering, using heuristics such as average digit to character ratio, punctuation to character ratio, average sentence length. This is then followed by the removal of duplicates across datasets using n-gram (n=5 in this work) matching, as there are potential overlaps between data sources. This preprocessing process ensures a high quality corpora of data, particularly important in resource-constrained environments.

Dataset Sources

Dataset	Description	Chars (before pre-processing)	Chars (after pre-processing)	Ratio
CulturaX	Cleaned, Enormous, and Public: The Multilingual Fuel to Democratize Large Language Models for 167 Languages.	1.3B	1.2B	65.0%
Glot500	A dataset of natural language data collected by putting together more than 150 existing mono-lingual and multilingual datasets together and crawling known multilingual websites.	1.3B	483.7M	27.1%
Gaparacrawl	Translation sentences crawl from the Internet	395.2M	106.3M	6.0%
Gawikipedia	Text from Irish Wikipedia	41.9M	23.4M	1.3%
Corpora Irish	The Leipzig Corpora Collection offers free online access to 136 monolingual dictionaries enriched with statistical information.	37.1	11.1M	0.6%
Total	-	3.1B	1.8B	100%

Load the dataset

You can load the dataset in Python using the datasets library:

from datasets import load_dataset

dataset = load_dataset("ReliableAI/Irish-Text-Collection")

Considerations for Using the Data

As this corpus is a collection of multiple datasets from Internet sources, it might still contain personal and sensitive information. Before using this dataset for any purpose, including training deep learning models, this should be to be taken into account.

Dataset Structure

{
    "text": ...,
    "meta": ...,
    "id": ...,
}

Citation

To cite this dataset, please use:

@misc{tran2024uccix,
      title={UCCIX: Irish-eXcellence Large Language Model}, 
      author={Khanh-Tung Tran and Barry O'Sullivan and Hoang D. Nguyen},
      year={2024},
      eprint={2405.13010},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Datasets:

ReliableAI
/

Irish-Text-Collection

You need to agree to share your contact information to access this dataset