rustemgareev/wortschatz

Leipzig Corpora Collection

Text corpora from Wortschatz Leipzig (Universität Leipzig): 16 corpora in 6 languages, 8,501,333 sentences, plus word lists, source URLs and co-occurrence tables.

Load

from datasets import load_dataset

ds = load_dataset("rustemgareev/wortschatz", "sentences")
print(ds["rus_news_2024_1M"][0])
# {'s_id': 1, 'sentence': '...'}

One config per table type, one split per corpus.

What's inside

config	columns	rows	note
sentences	s_id, sentence	8,501,333
words	w_id, word, word_ci, freq	6,828,545	word_ci is empty except in chv_community_2017
sources	so_id, source, date	3,176,950	source = URL, date = as scraped
co_n	w1_id, w2_id, freq, sig	4,737,462	co-occurring words as immediate neighbours
co_s	w1_id, w2_id, freq, sig	28,374,188	co-occurring words in the same sentence
inv_w	w_id, s_id, pos	144,585,828	word position in the sentence; empty for rus_web_2002 and rus_wikipedia_2010
inv_so	so_id, s_id	8,501,333	links sentences to sources
meta	run, attribute, value	140	only 8 of 16 corpora

Corpora

split	original name	language	domain	year	sentences
bak_news_2017_100K	bak_news_2017_100K	Bashkir	news	2017	100,000
bul_news_2022_100K	bul_news_2022_100K	Bulgarian	news	2022	100,000
chv_community_2017	chv_community_2017	Chuvash	community	2017	71,333
hin_news_2022_100K	hin_news_2022_100K	Hindi	news	2022	100,000
ind_news_2024_100K	ind_news_2024_100K	Indonesian	news	2024	100,000
ind_news_2024_10K	ind_news_2024_10K	Indonesian	news	2024	10,000
rus_md_web_2016_1M	rus-md_web_2016_1M	Russian (Moldova)	web	2016	1,000,000
rus_ru_web_public_2019_10K	rus-ru_web-public_2019_10K	Russian (Russia)	web	2019	10,000
rus_ru_web_public_2019_1M	rus-ru_web-public_2019_1M	Russian (Russia)	web	2019	1,000,000
rus_ua_web_2015_1M	rus-ua_web_2015_1M	Russian (Ukraine)	web	2015	1,000,000
rus_newscrawl_public_2018_1M	rus_newscrawl-public_2018_1M	Russian	newscrawl	2018	1,000,000
rus_news_2024_10K	rus_news_2024_10K	Russian	news	2024	10,000
rus_news_2024_1M	rus_news_2024_1M	Russian	news	2024	1,000,000
rus_web_2002_1M_text	rus_web_2002_1M-text	Russian	web	2002	1,000,000
rus_wikipedia_2010_1M_text	rus_wikipedia_2010_1M-text	Russian	Wikipedia	2010	1,000,000
rus_wikipedia_2021_1M	rus_wikipedia_2021_1M	Russian	Wikipedia	2021	1,000,000

Notes

• Split names use underscores (rus_ua_web_2015_1M) because HuggingFace does not allow hyphens; file names keep the original names.
• Files are tab-separated without header rows; column names come from the dataset configuration.
• Empty word_ci / pos columns were added so that all splits of a config share the same schema.

License

© Universität Leipzig, Wortschatz Leipzig. According to the Terms of Usage, the downloadable text corpora are published under CC BY 4.0.