By completing the form below, you acknowledge that the provided data is offered as is. Although we anticipate no problems, you accept full responsibility for any repercussions resulting from the use of this data. Furthermore, you agree that the data must not be utilized for malicious or harmful purposes towards humanity.

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:

You may need to follow these instructions to setup authentication before downloading the dataset: https://huggingface.co/docs/huggingface_hub/quick-start#login

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

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 (%)
0	en	English	3,241,065,682	2,846,970,578,793	45.13
1	ru	Russian	799,310,908	737,201,800,363	11.69
2	es	Spanish	450,937,645	373,845,662,394	5.93
3	de	German	420,017,484	357,030,348,021	5.66
4	fr	French	363,754,348	319,332,674,695	5.06
5	zh	Chinese	218,624,604	227,055,380,882	3.60
6	it	Italian	211,309,922	165,446,410,843	2.62
7	pt	Portuguese	190,289,658	136,941,763,923	2.17
8	pl	Polish	142,167,217	117,269,087,143	1.86
9	ja	Japanese	111,188,475	107,873,841,351	1.71
10	nl	Dutch	117,392,666	80,032,209,900	1.27
11	ar	Arabic	74,027,952	69,354,335,076	1.10
12	tr	Turkish	94,207,460	64,292,787,164	1.02
13	cs	Czech	65,350,564	56,910,486,745	0.90
14	vi	Vietnamese	57,606,341	55,380,123,774	0.88
15	fa	Persian	59,531,144	45,947,657,495	0.73
16	hu	Hungarian	44,132,152	43,417,981,714	0.69
17	el	Greek	51,430,226	43,147,590,757	0.68
18	ro	Romanian	40,325,424	39,647,954,768	0.63
19	sv	Swedish	49,709,189	38,486,181,494	0.61
20	uk	Ukrainian	44,740,545	38,226,128,686	0.61
21	fi	Finnish	30,467,667	28,925,009,180	0.46
22	ko	Korean	20,557,310	24,765,448,392	0.39
23	da	Danish	25,429,808	22,921,651,314	0.36
24	bg	Bulgarian	24,131,819	22,917,954,776	0.36
25	no	Norwegian	18,907,310	18,426,628,868	0.29
26	hi	Hindi	19,665,355	16,791,362,871	0.27
27	sk	Slovak	18,582,517	16,442,669,076	0.26
28	th	Thai	20,960,550	15,717,374,014	0.25
29	lt	Lithuanian	13,339,785	14,247,110,836	0.23
30	ca	Catalan	15,531,777	12,530,288,006	0.20
31	id	Indonesian	23,251,368	12,062,966,061	0.19
32	bn	Bangla	12,436,596	9,572,929,804	0.15
33	et	Estonian	8,004,753	8,805,656,165	0.14
34	sl	Slovenian	7,335,378	8,007,587,522	0.13
35	lv	Latvian	7,136,587	7,845,180,319	0.12
36	he	Hebrew	4,653,979	4,937,152,096	0.08
37	sr	Serbian	4,053,166	4,619,482,725	0.07
38	ta	Tamil	4,728,460	4,378,078,610	0.07
39	sq	Albanian	5,205,579	3,648,893,215	0.06
40	az	Azerbaijani	5,084,505	3,513,351,967	0.06
41	kk	Kazakh	2,733,982	2,802,485,195	0.04
42	ur	Urdu	2,757,279	2,703,052,627	0.04
43	ka	Georgian	3,120,321	2,617,625,564	0.04
44	hy	Armenian	2,964,488	2,395,179,284	0.04
45	is	Icelandic	2,373,560	2,350,592,857	0.04
46	ml	Malayalam	2,693,052	2,100,556,809	0.03
47	ne	Nepali	3,124,040	2,061,601,961	0.03
48	mk	Macedonian	2,762,807	2,003,302,006	0.03
49	mr	Marathi	2,266,588	1,955,227,796	0.03
50	mn	Mongolian	1,928,828	1,850,667,656	0.03
51	be	Belarusian	1,643,486	1,791,473,041	0.03
52	te	Telugu	1,822,865	1,566,972,146	0.02
53	gl	Galician	1,785,963	1,382,539,693	0.02
54	eu	Basque	1,598,822	1,262,066,759	0.02
55	kn	Kannada	1,352,142	1,242,285,201	0.02
56	gu	Gujarati	1,162,878	1,131,730,537	0.02
57	af	Afrikaans	826,519	1,119,009,767	0.02
58	my	Burmese	865,575	882,606,546	0.01
59	si	Sinhala	753,655	880,289,097	0.01
60	eo	Esperanto	460,088	803,948,528	0.01
61	km	Khmer	1,013,181	746,664,132	0.01
62	pa	Punjabi	646,987	727,546,145	0.01
63	cy	Welsh	549,955	576,743,162	0.01
64	ky	Kyrgyz	570,922	501,442,620	0.01
65	ga	Irish	304,251	376,947,935	0.01
66	ps	Pashto	376,914	363,007,770	0.01
67	am	Amharic	243,349	358,206,762	0.01
68	ku	Kurdish	295,314	302,990,910	0.00
69	tl	Filipino	348,453	242,086,456	0.00
70	yi	Yiddish	141,156	217,584,643	0.00
71	lo	Lao	217,842	168,256,876	0.00
72	fy	Western Frisian	223,268	167,193,111	0.00
73	sd	Sindhi	109,162	147,487,058	0.00
74	mg	Malagasy	115,910	142,685,412	0.00
75	or	Odia	153,461	100,323,213	0.00
76	as	Assamese	52,627	83,787,896	0.00
77	ug	Uyghur	47,035	77,677,306	0.00
78	uz	Uzbek	87,219	75,250,787	0.00
79	la	Latin	48,968	44,176,580	0.00
80	hr	Croatian	460,690	40,796,811	0.00
81	sw	Swahili	66,506	30,708,309	0.00
82	ms	Malay	238,151	19,375,976	0.00
83	br	Breton	43,765	13,987,037	0.00
84	sa	Sanskrit	16,290	13,561,367	0.00
85	gd	Scottish Gaelic	8,408	4,796,485	0.00
86	su	Sundanese	1,554	1,308,460	0.00
87	jv	Javanese	2,058	625,429	0.00
88	tg	Tajik	483,835	-	-
89	ceb	Cebuano	263,890	-	-
90	tt	Tatar	218,102	-	-
91	ckb	Central Kurdish	172,035	-	-
92	lb	Luxembourgish	165,891	-	-
93	mt	Maltese	151,320	-	-
94	nn	Norwegian Nynorsk	126,083	-	-
95	qu	Quechua	1,202	72,101	0.00
96	ba	Bashkir	71,957	-	-
97	arz	Egyptian Arabic	71,625	-	-
98	dv	Divehi	66,702	-	-
99	bo	Tibetan	54,185	-	-
100	sh	Serbian (Latin)	45,619	-	-
101	yo	Yoruba	192	42,943	0.00
102	bs	Bosnian	1,237	39,768	0.00
103	azb	South Azerbaijani	29,833	-	-
104	ht	Haitian Creole	12	26,183	0.00
105	war	Waray	23,687	-	-
106	cv	Chuvash	22,570	-	-
107	sah	Sakha	22,141	-	-
108	li	Limburgish	206	18,532	0.00
109	ce	Chechen	17,322	-	-
110	pnb	Western Panjabi	15,625	-	-
111	nds	Low German	15,139	-	-
112	tk	Turkmen	14,393	-	-
113	gn	Guarani	103	12,708	0.00
114	oc	Occitan	10,556	-	-
115	xmf	Mingrelian	9,706	-	-
116	ast	Asturian	9,002	-	-
117	os	Ossetic	8,596	-	-
118	mhr	Eastern Mari	7,883	-	-
119	pms	Piedmontese	7,566	-	-
120	als[*]	Swiss German	6,936	-	-
121	vo	Volapük	6,621	-	-
122	so	Somali	39	6,053	0.00
123	bpy	Bishnupriya	5,087	-	-
124	new	Newari	4,344	-	-
125	hsb	Upper Sorbian	4,244	-	-
126	lmo	Lombard	3,530	-	-
127	an	Aragonese	2,746	-	-
128	ilo	Iloko	2,328	-	-
129	mzn	Mazanderani	1,914	-	-
130	lez	Lezghian	1,806	-	-
131	rm	Romansh	30	1,769	0.00
132	krc	Karachay-Balkar	1,745	-	-
133	min	Minangkabau	1,429	-	-
134	kv	Komi	1,396	-	-
135	wa	Walloon	1,383	-	-
136	jbo	Lojban	1,349	-	-
137	io	Ido	1,144	-	-
138	mrj	Western Mari	1,056	-	-
139	gom	Goan Konkani	721	-	-
140	ia	Interlingua	613	-	-
141	av	Avaric	438	-	-
142	bh	Bihari languages	265	-	-
143	wuu	Wu Chinese	222	-	-
144	nah	Nahuatl languages	131	-	-
145	vec	Venetian	113	-	-
146	bxr	Russia Buriat	100	-	-
147	kw	Cornish	94	-	-
148	mai	Maithili	93	-	-
149	eml[*]	Emiliano-Romagnol	91	-	-
150	dsb	Lower Sorbian	59	-	-
151	xal	Kalmyk	51	-	-
152	lrc	Northern Luri	43	-	-
153	nap	Neapolitan	31	-	-
154	tyv	Tuvinian	23	-	-
155	scn	Sicilian	21	-	-
156	frr	Northern Frisian	11	-	-
157	mwl	Mirandese	9	-	-
158	myv	Erzya	4	-	-
159	ie	Interlingue	4	-	-
160	pam	Pampanga	4	-	-
161	bar	Bavarian	3	-	-
162	yue	Yue Chinese	3	-	-
163	cbk	Chavacano	2	-	-
164	bcl	Central Bikol	1	-	-
165	vls	West Flemish	1	-	-
166	rue	Rusyn	1	-	-

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.

Acknowledgements

We would like to extend our sincere thanks to Google Cloud for providing the TPU resources that made this project possible. Their support has been invaluable in enabling our team to run evaluations on our dataset efficiently.

Citation

To cite CulturaX, please use:

@inproceedings{nguyen-etal-2024-culturax,
    title = "{C}ultura{X}: A Cleaned, Enormous, and Multilingual Dataset for Large Language Models in 167 Languages",
    author = "Nguyen, Thuat  and
      Nguyen, Chien Van  and
      Lai, Viet Dac  and
      Man, Hieu  and
      Ngo, Nghia Trung  and
      Dernoncourt, Franck  and
      Rossi, Ryan A.  and
      Nguyen, Thien Huu",
    editor = "Calzolari, Nicoletta  and
      Kan, Min-Yen  and
      Hoste, Veronique  and
      Lenci, Alessandro  and
      Sakti, Sakriani  and
      Xue, Nianwen",
    booktitle = "Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)",
    month = may,
    year = "2024",
    address = "Torino, Italia",
    publisher = "ELRA and ICCL",
    url = "https://aclanthology.org/2024.lrec-main.377",
    pages = "4226--4237",
    abstract = "Extensive training datasets represent one of the important factors for the impressive learning capabilities of large language models (LLMs). However, these training datasets for current LLMs, especially the recent state-of-the-art models, are often not fully disclosed. Creating training data for high-performing LLMs involves extensive cleaning and deduplication to ensure the necessary level of quality. The lack of transparency for training data has thus hampered research on attributing and addressing hallucination and bias issues in LLMs, hindering replication efforts and further advancements in the community. These challenges become even more pronounced in multilingual learning scenarios, where the available multilingual text datasets are often inadequately collected and cleaned. Consequently, there is a lack of open-source and readily usable dataset to effectively train LLMs in multiple languages. To overcome this issue, we present CulturaX, a substantial multilingual dataset with 6.3 trillion tokens in 167 languages, tailored for 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. CulturaX is released in Hugging Face facilitate research and advancements in multilingual LLMs: https://huggingface.co/datasets/uonlp/CulturaX.",
}

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.