CulturaX

Cleaned, Enormous, and Public: The Multilingual Fuel to Democratize Large Language Models for 167 Languages

Dataset Summary

We present CulturaX, a substantial multilingual dataset with 6.3 trillion tokens in 167 languages, tailored for large language model (LLM) development. Our dataset undergoes meticulous cleaning and deduplication through a rigorous pipeline of multiple stages to accomplish the best quality for model training, including language identification, URL-based filtering, metric-based cleaning, document refinement, and data deduplication. We employ MinHash at document level to achieve fuzzy deduplication for the datasets in different languages. Our data cleaning framework includes diverse criteria and threshold selections, guided by extensive data samples, ensuring comprehensive noise filtering in various aspects. CulturaX is fully released to the public in HuggingFace to facilitate research and advancements in multilingual LLMs.

Our dataset combines the most recent iteration of mC4 (version 3.1.0) [1] with all accessible OSCAR corpora up to the present year, including 20.19, 21.09, 22.01, and 23.01 [2]. After deep cleaning and deduplication, CulturaX involves 16TB data in the parquet format (expanding to 27TB when unpacked). More than a half of our dataset is dedicated to non-English languages to significantly boost the data size and enhance the feasibility of training models in multilingual scenarios.

To obtain perplexity scores for data cleaning, we train a SentencePiece tokenizer and 5-gram Kneser-Ney language models as provided in the KenLM library [3] using the 20230501 dumps of Wikipedia. Our KenLM models are also released in HuggingFace: https://huggingface.co/uonlp/kenlm.

Details for the dataset can be found in our technical paper: https://arxiv.org/abs/2309.09400

You can download the dataset using Hugging Face datasets:

from datasets import load_dataset
ds = load_dataset("uonlp/CulturaX", "en")

Languages

The supported languages and statistics for our dataset can be found below:

(Note that the language code als and eml refer to gsw and x-eml in the OSCAR-2301 dataset.)

	Code	Language	# Documents	# Tokens	# Tokens (%)
...	...	...	...	...	...
10	vi	Vietnamese	102,411,180	98,453,464,077	1.56
...	...	...	...	...	...

Dataset Structure

{
    "text": ...,
    "timestamp": ...,
    "url": ...,
    "source": "mc4" | "OSCAR-xxxx",
}

Considerations for Using the Data

As CulturaX is the cleaned version of the mC4 and OSCAR datasets, which were both extracted from CommonCrawl, personal and sensitive information might still contain personal and sensitive information. This must be considered prior to using this dataset for any purpose, such as training deep learning models, etc.

License Information

The licence terms for CulturaX strictly follows those of mC4 and OSCAR. Please refer to both below licenses when using this dataset.

Citation

To cite CulturaX, please use:

@misc{nguyen2023culturax,
      title={CulturaX: A Cleaned, Enormous, and Multilingual Dataset for Large Language Models in 167 Languages}, 
      author={Thuat Nguyen and Chien Van Nguyen and Viet Dac Lai and Hieu Man and Nghia Trung Ngo and Franck Dernoncourt and Ryan A. Rossi and Thien Huu Nguyen},
      year={2023},
      eprint={2309.09400},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Reference

[1] Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mT5: A massively multilingual pre-trained text-to-text transformer. In NAACL 2021. https://huggingface.co/datasets/mc4

[2] Pedro Javier Ortiz Suárez, Benoît Sagot, and Laurent Romary. 2019. Asynchronous pipelines for processing huge corpora on medium to low resource infrastructures. In Proceedings of the Workshop on Challenges in the Management of Large Corpora (CMLC- 7) 2019. https://oscar-project.org/

[3] KenLM: Faster and smaller language model queries. In Proceedings of the Sixth Workshop on Statistical Machine Translation, 2011.

Datasets:

vietgpt
/

CulturaX

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