Datasets:

VSSA-SDSA
/

LT_AI_BLKT

language:
  - lt
pretty_name: General Lithuanian Corpus
tags:
  - text
  - parquet
  - corpus
  - lithuanian
license: openrail
size_categories:
  - 1M<n<10M
task_categories:
  - text-generation
  - summarization
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/*.parquet

Model Card for LT_AI_BLKT (EN) / LT_AI_BLKT modelio kortelė (LT)

Table of Contents / Turinys

Description (EN) / Aprašas (LT)
Dataset summary(EN)
Main columns (EN)
Data composition (EN)
Distribution of text types (EN)
Sources (EN)
Time periods (EN)
Licensing (EN) / Licencija (LT))
Intended use (EN) / Numatyti naudojimo atvejai (LT)
Restrictions] (EN) / Apribojimai (LT)
Limitations and bias (EN) / Ribotumai ir šališkumas (LT)
Citation (EN) / Citavimas (LT)

English

Description

The General Lithuanian Language Corpus (hereinafter BLKT) was developed for the State Digital Solutions Agency (SDSA) as part of the project Creation of the General Lithuanian Language Corpus and Vectorized Models (project code: 02-103-P-0001). The corpus was developed by a consortium of four partners: Vytautas Magnus University (VMU), UAB “Tilde informacinės technologijos”, UAB “Neurotechnology”, and MB “Krilas”.

The project is implemented under Progress Measure No. 05-002-01-07-08 of the 2021–2030 Lithuanian Republic State Digitalization Development Programme, Developing technological solutions and tools that enable safe and convenient use of services, and fulfils the requirements of the activity Development of linguistic resources for the needs of artificial intelligence technology solutions.

Dataset summary

Number of Parquet files: 25
Total rows (texts): 8,438,155
Total file size: 17.79 GB
Total alpha-word count: 3,941,476,219

Main columns

alpha_word_count: int64
author: string
document_subtype: string
document_type: string
id: string
language: string
license: string
period: string
publication_date: string
record_created: string
source_file: string
source_id: string
source_name: string
text: string
text_char_count: int64
title: string
url: string

Data composition

All BLKT texts are in Lithuanian and are human-written. The corpus is divided into five main text types (document_type): fiction (gro), non-fiction (neg), media (zin), speech (sak), and documents (dok). These are further divided into text subtypes (document_subtype), as shown below.

Type	Description	Subtype	Description
zin	Media	port	News portals
zin	Media	prdk	Newspapers and journals
sak	Speech	sste	Parliament transcriptions
sak	Speech	skit	Other speech transcriptions
neg	Non-fiction	intt	Internet texts
neg	Non-fiction	moks	Scientific papers and books
neg	Non-fiction	kiti	Other texts
gro	Fiction	proz	Prose
gro	Fiction	poez	Poetry
dok	Documents	dkes	EU documents
dok	Documents	dklt	Lithuanian legal documents
dok	Documents	dkad	Administrative documents

Distribution of text types

Text type	Word count	Share	Text count
Documents	1,509,347,610	38.29%	1,212,131
Fiction	11,767,755	0.30%	440
Non-fiction	279,251,467	7.08%	832,256
Speech	80,483,739	2.04%	4,563
Media	2,060,625,648	52.28%	6,388,765
Total	3,941,476,219	100.00%	8,438,155

Sources

The corpus was compiled from 36 different sources, including news portals, legal and administrative documents, scientific publications, internet texts, speech transcriptions, and fiction. All source materials were included on the basis of the necessary permissions, licences, or other lawful grounds for use, in accordance with applicable copyright and data protection requirements.

Time periods

The texts in the corpus were published across four broad periods.

Time span	Period	Word count	Share	Text count
1922–1940	1	11,226,438	0.28%	290
1941–1990	2	10,440,807	0.26%	4,288
1990–2004	3	229,775,041	5.83%	684,357
2008–2026	4	3,690,033,933	93.62%	7,749,220
	Total	3,941,476,219	100.00%	8,438,155

Licensing

This dataset is distributed under the NewGenLTU OpenRAIL-D licence.

This licence was specifically created by a team of researchers at Vytautas Magnus University for the publication of datasets produced in the course of research and development projects funded by the New Generation Lithuania plan. The licence aims to support open and responsible downstream use while ensuring compliance with applicable requirements related to copyright, personal data protection, and the EU AI Act.

Please see the accompanying LICENSE.txt file in this repository for the full licence text.

Intended use

This dataset may be used for a range of Lithuanian-language NLP and AI tasks, including:

text generation,
summarisation,
language modelling,
grammar and style correction,
semantic search,
text analysis,
virtual assistants,
other language technology applications.

Restrictions

Under the NewGenLTU OpenRAIL-D licence, users must not use the dataset for:

discrimination,
the development of weapons or military applications,
automated decision-making affecting people,
disinformation,
privacy infringement,
medical or health advice,
creation of malware,
harassment,
dishonest research,
collection, extraction, or reconstruction of personal data.

Please refer to the accompanying LICENSE.txt file for the full and authoritative wording.

Limitations and bias

The developers made substantial efforts to clean the dataset and reduce noise, OCR errors, and duplicates. However, users should be aware of the following limitations:

proportion of spelling errors: 0.1%,
proportion of foreign words: 0.07%,
texts from periods 1–2 (0.54% of the corpus) may contain archaic or no longer commonly used vocabulary and expressions,
the corpus is strongly dominated by news portal texts (52%) and documents (38%), which may bias downstream models toward these registers and domains.

Citation

Please cite the dataset as:

State Digital Solutions Agency. 2026. General Lithuanian Language Corpus. Hugging Face. https://huggingface.co/datasets/VSSA-SDSA/LT_AI_BLKT

BibTeX:

@dataset{gllc_2026,
  author    = {{State Digital Solutions Agency}},
  title     = {General Lithuanian Language Corpus},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/VSSA-SDSA/LT_AI_BLKT},
  note      = {Developed by Vytautas Magnus University, UAB "Tilde informacinės technologijos", UAB "Neurotechnology", and MB "Krilas".}
}

Lietuviškai

Aprašas

Bendrasis lietuvių kalbos tekstynas (toliau – BLKT) sukurtas Valstybės skaitmeninių sprendimų agentūrai (VSSA) įgyvendinant projektą Bendrojo lietuvių kalbos tekstyno ir vektorizuotų modelių sukūrimas (projekto kodas: 02-103-P-0001). Tekstyną sukūrė keturių partnerių konsorciumas: Vytauto Didžiojo universitetas (VDU), UAB „Tilde informacinės technologijos“, UAB „Neurotechnology“ ir MB „Krilas“.

Tekstyną sudaro lietuvių kalba parašyti žmogaus sukurti tekstai, suskirstyti į penkis pagrindinius tipus: grožinius, negrožinius, žiniasklaidos, sakytinės kalbos ir dokumentų tekstus. Duomenų rinkinį sudaro 8 438 155 tekstai, kurių bendras alfabetinių žodžių skaičius siekia 3 941 476 219. Duomenys pateikiami 25 Parquet failuose, kurių bendras dydis yra 17,79 GB.

Licencija

Šis duomenų rinkinys platinamas pagal NewGenLTU OpenRAIL-D licenciją. Visas licencijos tekstas pateiktas šiame saugykloje esančiame faile LICENSE.txt.

Numatyti naudojimo atvejai

Duomenų rinkinys gali būti naudojamas lietuvių kalbos technologijų užduotims, įskaitant:

teksto generavimą,
apibendrinimą,
gramatikos ir stiliaus taisymą,
semantinę paiešką,
teksto analizę,
virtualiuosius asistentus.

Apribojimai

Pagal NewGenLTU OpenRAIL-D licenciją šio duomenų rinkinio negalima naudoti diskriminacijai, ginklų ar karinių sprendimų kūrimui, automatizuotam sprendimų priėmimui, dezinformacijai, privatumo pažeidimams, medicininiams ar sveikatos patarimams, kenkėjiškų programų kūrimui, priekabiavimui, nesąžiningiems tyrimams ar asmens duomenų rinkimui, išgavimui arba atkūrimui.

Ribotumai

Nors tekstynas buvo kruopščiai valytas, naudotojai turėtų atsižvelgti į tai, kad:

rašybos klaidų dalis sudaro 0,1%,
kitų kalbų žodžių dalis sudaro 0,07%,
ankstyvesnių laikotarpių tekstuose gali pasitaikyti senesnės leksikos,
tekstyne vyrauja naujienų portalų ir dokumentų tekstai, todėl iš jo mokomi modeliai gali būti šališki šių registrų atžvilgiu.

Citavimas

Prašau cituokite šį rinkinį taip:

Valstybės skaitmeninių sprendimų agentūra. 2026. Bendrasis lietuvių kalbos tekstynas. Hugging Face. https://huggingface.co/datasets/VSSA-SDSA/LT_AI_BLKT .

BibTeX:

@dataset{blkt_2026,
  author    = {{Valstybės skaitmeninių sprendimų agentūra}},
  title     = {Bendrasis lietuvių kalbos tekstynas},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/VSSA-SDSA/LT_AI_BLKT},
  note      = {Sukurta Vytauto Didžiojo universiteto, UAB "Tilde informacinės technologijos", UAB "Neurotechnology" ir MB "Krilas".}
}