CulturaX: A High-Quality, Multilingual Dataset for LLMs - Data Analysis and Experiments

cover
28 Aug 2024

Authors:

(1) Thuat Nguyen, Dept. of Computer Science, University of Oregon, OR, USA;

(2) Chien Van Nguyen, Dept. of Computer Science, University of Oregon, OR, USA;

(3) Viet Dac Lai, Dept. of Computer Science, University of Oregon, OR, USA;

(4) Hieu Man, Dept. of Computer Science, University of Oregon, OR, USA;

(5) Nghia Trung Ngo, Dept. of Computer Science, University of Oregon, OR, USA;

(6) Franck Dernoncourt, Adobe Research, USA;

(7) Ryan A. Rossi, Adobe Research, USA;

(8) Thien Huu Nguyen, Dept. of Computer Science, University of Oregon, OR, USA.

Abstract and Introduction

Multilingual Dataset Creation

Data Analysis and Experiments

Related Work

Conclusion and References

3 Data Analysis and Experiments

After completing all the cleaning and deduplication steps, our ultimate dataset comprises 6.3 trillion tokens spanning 167 languages. Table 1 provides an overview of the number of documents and tokens for the top 42 languages in CulturaX following each processing stage. As can be seen, our datacleaning pipeline can substantially reduce the number of documents in the original mC4 and OSCAR datasets for each language. The total number of removed documents accounts for 46.48% of our initial documents, suggesting the the effectiveness of our approaches to filter noisy information for multilingual datasets.

This paper is available on arxiv under CC BY 4.0 DEED license.