Viet-Mistral/CulturaY

CulturaY: A Large Cleaned Multilingual Dataset of 75 Languages

Dataset Summary

From the team that brought you CulturaX, we present CulturaY, another substantial multilingual dataset of 15TB (uncompressed)/3TB (zstd-compressed) that applies the same dataset cleaning methodology to the HPLT v1.1 dataset. Please note that HPLT v1.2 has also been released and is an alternative verison with different cleaning methodolgies. This data was used in part to train our SOTA Vietnamese model: Vistral-7B-Chat.

Our annotations and arrangements are licensed under CC-BY-4.0, and we make the data available for fair use machine learning research.
But we make no claims as to the underlying copyrights of the work. This data was copied from the HPLT project, which in turn used the data from Common Crawl and the Internet Archive.

Acknowledgement

We thank our collaborators at UONLP - The Natural Language Processing Group at the University of Oregon, and the computing resources of the managers of the Karolina Supercomputers. We also thank our friends at TurkuNLP for their support.

Data Breakdown:

There are 75 langauges, with the following breakdown:

	Code	Language	# Documents	# Documents (%)	Size (GB)
0	en	English	523,235,685	43.84	1244.39
1	zh	Chinese	172,023,436	14.41	290.91
2	ru	Russian	59,185,035	4.96	424.55
3	es	Spanish	49,193,764	4.12	116.20
4	de	German	35,204,652	2.95	78.32
5	fr	French	33,063,792	2.77	69.66
6	ja	Japanese	27,641,765	2.32	74.71
7	ko	Korean	26,925,013	2.26	25.50
8	it	Italian	22,396,067	1.88	48.30
9	pt	Portuguese	18,367,640	1.54	39.09
10	th	Thai	16,330,227	1.37	32.09
11	da	Danish	13,547,169	1.13	18.40
12	sv	Swedish	13,049,359	1.09	19.29
13	tr	Turkish	12,659,104	1.06	29.14
14	nl	Dutch	12,454,669	1.04	22.58
15	pl	Polish	12,054,997	1.01	27.09
16	hu	Hungarian	11,939,984	1.00	17.63
17	ro	Romanian	11,578,945	0.97	18.57
18	hbs	Serbo-Croatian	8,880,450	0.74	14.65
19	id	Indonesian	8,473,141	0.71	16.23
20	bg	Bulgarian	6,698,866	0.56	18.63
21	el	Greek	6,674,496	0.56	29.61
22	ar	Arabic	6,427,386	0.54	28.04
23	nb	Norwegian Bokmål	5,925,942	0.50	10.14
24	fi	Finnish	5,379,100	0.45	10.08
25	he	Hebrew	5,320,279	0.45	12.06
26	uk	Ukrainian	5,311,749	0.45	31.55
27	cs	Czech	5,248,678	0.44	12.83
28	fa	Persian	5,111,868	0.43	26.23
29	ms	Malay	4,888,894	0.41	9.09
30	sk	Slovak	4,758,917	0.40	5.50
31	ca	Catalan	4,552,579	0.38	7.96
32	vi	Vietnamese	4,493,567	0.38	16.95
33	hi	Hindi	4,200,330	0.35	11.56
34	bn	Bangla	2,785,980	0.23	4.76
35	lt	Lithuanian	2,509,788	0.21	3.83
36	sl	Slovenian	2,252,359	0.19	3.21
37	la	Latin	2,147,688	0.18	1.42
38	et	Estonian	1,754,719	0.15	2.88
39	az	Azerbaijani	1,554,357	0.13	1.95
40	lv	Latvian	1,469,245	0.12	2.19
41	ur	Urdu	1,251,414	0.10	2.84
42	ta	Tamil	1,128,321	0.09	7.21
43	gl	Galician	1,101,337	0.09	1.31
44	sq	Albanian	1,081,763	0.09	1.73
45	ne	Nepali	860,657	0.07	1.91
46	mk	Macedonian	641,111	0.05	1.61
47	af	Afrikaans	636,976	0.05	0.77
48	tl	Filipino	575,221	0.05	1.09
49	sw	Swahili	571,247	0.05	0.60
50	eu	Basque	559,194	0.05	0.67
51	is	Icelandic	529,777	0.04	0.81
52	ka	Georgian	524,645	0.04	1.48
53	hy	Armenian	519,060	0.04	1.46
54	my	Burmese	513,729	0.04	1.91
55	nn	Norwegian Nynorsk	509,287	0.04	0.49
56	ml	Malayalam	487,912	0.04	2.02
57	mn	Mongolian	448,211	0.04	1.79
58	be	Belarusian	426,194	0.04	1.48
59	uz	Uzbek	423,865	0.04	1.19
60	mr	Marathi	398,138	0.03	1.28
61	si	Sinhala	337,785	0.03	1.55
62	te	Telugu	279,240	0.02	1.00
63	kk	Kazakh	274,770	0.02	1.07
64	mt	Maltese	265,605	0.02	0.90
65	so	Somali	261,100	0.02	0.24
66	gu	Gujarati	242,074	0.02	0.74
67	kn	Kannada	231,260	0.02	0.71
68	cy	Welsh	179,157	0.02	0.20
69	ga	Irish	134,796	0.01	0.15
70	tt	Tatar	131,731	0.01	0.41
71	pa	Punjabi	119,686	0.01	0.29
72	eo	Esperanto	114,598	0.01	0.17
73	ps	Pashto	99,783	0.01	0.23
74	ky	Kyrgyz	86,551	0.01	0.31

Dataset structure

The dataset has a total of 6 columns, including:

• 2 columns text, url will be the two main columns in this dataset.
• the remaining columns id, document_lang, scores, langs belong to the original document in the HPLT V1.1 dataset, retained for debugging purposes. and will be removed in the future.

Therefore, when using, please only utilize the two columns text and url.

Process for Creating CulturaY

Firstly, to create CulturaY, we began with the HPLT dataset (version 1.1). This is also a notable difference between X and Y. While X was generated from cleaning data from Common Crawl (mC4, Oscar), Y was generated from cleaning raw data from the Internet Archive (HPLT). While Common Crawl is quite popular, data from the Internet Archive is less known and exploited, even though the data from both sources are similar. HPLT or CulturaY could be considered the first publicly released datasets originating from the Internet Archive. Using both CulturaX and CulturaY simultaneously will help your model have a more diverse source of data.

Our pipeline is built based on Bloom's data cleaning pipeline: evaluating each document in the dataset according to criteria such as document length, perplexity, bad words ratio, etc., and removing documents that do not perform well in any of these criteria. See our Blog for more details.

Citation

To cite CulturaY, please use:

@misc{nguyen2024culturay,
      title={CulturaY: A Large Cleaned Multilingual Dataset of 75 Languages}, 
      author={Thuat Nguyen, Huu Nguyen and Thien Nguyen},
      year={2024},
}