README.md

metadata

pretty_name: SNeG-data-1.0
license: odc-by
language:
  - alt
  - atv
  - ba
  - be
  - bxr
  - ce
  - cv
  - de
  - en
  - kbd
  - kca
  - kjh
  - kk
  - koi
  - kpv
  - krc
  - krl
  - kum
  - ky
  - lez
  - lud
  - mdf
  - mhr
  - mns
  - mrj
  - myv
  - nog
  - olo
  - ru
  - sah
  - tt
  - tyv
  - udm
  - uz
  - vep
  - xal
task_categories:
  - text-generation
language_creators:
  - found
annotations_creators:
  - machine-generated
multilinguality:
  - multilingual
size_categories:
  - 1M<n<10M
tags:
  - text
  - corpus
  - multilingual
  - low-resource
  - pretraining
  - news
  - literature
  - folklore
  - cis
  - post-soviet
  - regional-languages
source_datasets:
  - d0rj/ilibrary
configs:
  - config_name: unfiltered
    data_files:
      - split: train
        path: data/unfiltered/train-*.parquet
  - config_name: filtered
    data_files:
      - split: train
        path: data/filtered/train-*.parquet

SNeG-data-1.0

English · Русский

SNeG-data-1.0 is a multilingual collection of news articles, literature, folklore, educational and reference materials, and other texts in languages of the CIS and the wider post-Soviet region, with an emphasis on low-resource languages.

Statistics

Both subsets contain a single shuffled train split. unfiltered contains the collected texts before cleaning; filtered contains the cleaned and deduplicated texts.

Subset	Records	Gemma 4 tokens
unfiltered	3,046,707	4,125,211,521
filtered	2,985,422	3,990,012,437

Token counts cover the text field, using the Gemma 4 tokenizer without special tokens. A record may be an article, a literary work, an excerpt or a sentence, depending on the source.

The distributions below use the declared source language (source_language); translations are counted with their source.

filtered: distribution by source language

Source language	Code	Records	Record share	Tokens	Token share
Russian	rus	17,299	0.5794%	789,172,812	19.7787%
Bashkir	bak	491,794	16.4732%	689,390,975	17.2779%
Tatar	tat	599,197	20.0708%	535,045,840	13.4096%
Kazakh	kaz	346,029	11.5906%	487,209,802	12.2107%
Uzbek	uzb	352,852	11.8192%	407,081,528	10.2025%
Belarusian	bel	234,752	7.8633%	395,883,536	9.9219%
Kyrgyz	kir	353,217	11.8314%	229,009,384	5.7396%
Sakha (Yakut)	sah	94,522	3.1661%	115,948,096	2.9060%
Kabardian	kbd	21,466	0.7190%	46,175,376	1.1573%
Chechen	che	41,209	1.3803%	44,161,085	1.1068%
Lezgi	lez	19,191	0.6428%	30,692,403	0.7692%
Chuvash	chv	50,393	1.6880%	30,672,857	0.7687%
Udmurt	udm	25,160	0.8428%	22,535,813	0.5648%
Tuvan	tyv	15,330	0.5135%	20,724,278	0.5194%
Khanty	kca	13,135	0.4400%	16,860,557	0.4226%
Mansi	mns	203,731	6.8242%	16,627,771	0.4167%
Khakas	kjh	11,351	0.3802%	15,449,350	0.3872%
Komi-Zyrian	kpv	11,984	0.4014%	15,308,759	0.3837%
Karachay-Balkar	krc	12,755	0.4272%	15,138,986	0.3794%
Southern Altai	alt	5,623	0.1883%	13,898,826	0.3483%
Buryat	bxr	9,265	0.3103%	12,051,140	0.3020%
Meadow Mari	mhr	12,954	0.4339%	10,156,359	0.2545%
Erzya	myv	7,679	0.2572%	8,314,047	0.2084%
Multilingual sources	mul	17,907	0.5998%	7,153,913	0.1793%
Livvi-Karelian	olo	5,257	0.1761%	5,121,126	0.1283%
Karelian	krl	5,120	0.1715%	4,328,175	0.1085%
Veps	vep	3,790	0.1270%	2,811,756	0.0705%
Kumyk	kum	254	0.0085%	1,136,101	0.0285%
Moksha	mdf	354	0.0119%	539,479	0.0135%
Kalmyk	xal	997	0.0334%	528,473	0.0132%
Ludian	lud	615	0.0206%	492,043	0.0123%
Hill Mari	mrj	204	0.0068%	177,974	0.0045%
Komi-Permyak	koi	11	0.0004%	148,487	0.0037%
Unspecified source language	und	4	0.0001%	30,925	0.0008%
Nogai	nog	13	0.0004%	19,998	0.0005%
Northern Altai	atv	8	0.0003%	14,407	0.0004%
Total		2,985,422	100.0000%	3,990,012,437	100.0000%

unfiltered: distribution by source language

Source language	Code	Records	Record share	Tokens	Token share
Russian	rus	18,801	0.6171%	809,266,734	19.6176%
Bashkir	bak	516,532	16.9538%	776,125,605	18.8142%
Tatar	tat	608,590	19.9753%	545,701,048	13.2284%
Kazakh	kaz	347,920	11.4195%	490,823,165	11.8981%
Uzbek	uzb	352,897	11.5829%	407,233,698	9.8718%
Belarusian	bel	234,906	7.7102%	396,201,630	9.6044%
Kyrgyz	kir	353,358	11.5980%	229,272,199	5.5578%
Sakha (Yakut)	sah	95,244	3.1261%	117,509,727	2.8486%
Kabardian	kbd	25,241	0.8285%	53,874,649	1.3060%
Chechen	che	41,238	1.3535%	44,196,457	1.0714%
Lezgi	lez	19,308	0.6337%	31,264,322	0.7579%
Chuvash	chv	50,400	1.6542%	30,705,760	0.7443%
Udmurt	udm	25,817	0.8474%	23,038,646	0.5585%
Tuvan	tyv	15,397	0.5054%	20,827,966	0.5049%
Khanty	kca	13,155	0.4318%	16,986,127	0.4118%
Mansi	mns	218,015	7.1558%	16,977,575	0.4116%
Komi-Zyrian	kpv	12,947	0.4250%	15,797,237	0.3829%
Karachay-Balkar	krc	13,394	0.4396%	15,782,106	0.3826%
Khakas	kjh	11,359	0.3728%	15,561,748	0.3772%
Southern Altai	alt	5,691	0.1868%	14,135,136	0.3427%
Buryat	bxr	9,271	0.3043%	12,093,101	0.2932%
Meadow Mari	mhr	12,994	0.4265%	10,193,703	0.2471%
Erzya	myv	8,375	0.2749%	8,720,583	0.2114%
Multilingual sources	mul	18,060	0.5928%	7,164,586	0.1737%
Livvi-Karelian	olo	6,005	0.1971%	5,365,392	0.1301%
Karelian	krl	5,499	0.1805%	4,464,484	0.1082%
Veps	vep	3,791	0.1244%	2,813,366	0.0682%
Kumyk	kum	254	0.0083%	1,136,118	0.0275%
Moksha	mdf	396	0.0130%	566,103	0.0137%
Kalmyk	xal	997	0.0327%	528,506	0.0128%
Ludian	lud	615	0.0202%	492,253	0.0119%
Hill Mari	mrj	204	0.0067%	177,974	0.0043%
Komi-Permyak	koi	11	0.0004%	148,487	0.0036%
Unspecified source language	und	4	0.0001%	30,925	0.0007%
Nogai	nog	13	0.0004%	19,998	0.0005%
Northern Altai	atv	8	0.0003%	14,407	0.0003%
Total		3,046,707	100.0000%	4,125,211,521	100.0000%

Filtering

• Text cleaning. Confirmed website templates, navigation elements, widgets, utility links and credits are removed. Line breaks, split words and whitespace are repaired.
• Quality filtering. Empty and clearly corrupted texts, error and access-challenge pages, unparsed binary documents and degenerate repetition are removed.
• Exact deduplication. Document duplicates are identified after Unicode NFC and whitespace normalization.
• Near-duplicate removal. MinHash over word 5-grams retrieves candidates, which are compared using their full texts within the same source-language group. Texts must contain at least 80 words, have a length ratio of at least 0.98 and identical numeric sequences. Removal requires either an identical complete normalized word sequence, or a matching title, Jaccard similarity of at least 0.98 and complete ordered containment of the shorter text in the retained text. Source-language groups with fewer than 2 million original tokens require identical complete word sequences.
• Language annotation. GlotLID v3, based on fastText, identifies document and paragraph languages. Context and verified publisher labels for individual texts or translation columns help resolve short or ambiguous passages.

Meaningful repetition in verse, refrains, quotations and parallel writing systems is preserved.

Records include source provenance, source category and language annotations. Source rights and license information, where available, is retained in rights_status and license_url.

License

The SNeG-data-1.0 database compilation is licensed under the ODC Attribution License (ODC-By) 1.0, to the extent of the compiler's rights in the selection and arrangement of the materials. The license permits personal, educational and commercial use of the database. Public use of the database, derivative databases or works produced from it requires attribution linking to SNeG-data-1.0 and ODC-By 1.0, as specified in the license.

Individual source texts remain subject to their respective rights holders' copyright.

Citation

@misc{sneg_data_2026,
  author       = {{d0rj}},
  title        = {{SNeG-data-1.0}},
  year         = {2026},
  howpublished = {Hugging Face},
  url          = {https://huggingface.co/datasets/d0rj/SNeG-data-1.0}
}