The Tatar Web Corpus is the largest open-source corpus of the Tatar language (Turkic family), containing 2,465,867 documents (approximately 251 million tokens) collected from publicly available web sources. It covers news portals, social media, blogs, literary websites, and other domains. The corpus underwent soft deduplication to remove exact duplicates while preserving linguistic diversity.

Curated by: TatarNLPWorld team (Mullosharaf K. Arabov and contributors)
Language(s) (NLP): Tatar (tt)
License: other – see Licensing & Legal Notice below.

Licensing & Legal Notice

This dataset follows the practice established by large web-crawled corpora such as HPLT and OSCAR:

Original source texts (web pages, articles, posts) remain the property of their respective authors and publishers. They are not owned by the TatarNLPWorld team and are not covered by the CC BY-SA 4.0 license.
The structured compilation, metadata, and any original annotations created by TatarNLPWorld are licensed under CC BY-SA 4.0.
Users are solely responsible for ensuring their use of the underlying texts complies with applicable copyright law. For commercial use of verbatim excerpts, permission from the original copyright holders may be required.
A notice-and-takedown procedure is in place: rights holders can request removal of specific content by contacting the dataset maintainers (see Dataset Card Contact). We commit to responding within 14 business days and removing disputed content in the next release.

Dataset Sources

Repository: https://huggingface.co/datasets/TatarNLPWorld/tatar-web-corpus-v3
Papers:
- Arabov, M. K., Gilmullin, R. A., & Burnashev, R. A. (2026). Combining Classical and Transformer-based Approaches for Text Classification and Topic Modeling of the Tatar Language. In 2026 IEEE Ural-Siberian Conference on Biomedical Engineering, Radioelectronics and Information Technology (USBEREIT) (pp. 1–4). IEEE. https://doi.org/10.1109/USBEREIT70063.2026.11580632
- Burnashev, R. A., Gilmullin, R. A., & Arabov, M. K. (2026). Parameter-efficient fine-tuning of large language models for Tatar language: a comparative study of LoRA and QLoRA. In 2026 IEEE Ural-Siberian Conference on Biomedical Engineering, Radioelectronics and Information Technology (USBEREIT) (pp. 1–4). IEEE. https://doi.org/10.1109/USBEREIT70063.2026.11580544
- Arabov, M. K., Gainullin, R. N., & Khusaenov, A. I. (2026). Modelling Subword Embeddings of Low-Resource Languages for the Digitalisation of Industry 4.0 Industrial Systems. In 2026 International Russian Smart Industry Conference (SmartIndustryCon) (pp. 74–79). IEEE. https://doi.org/10.1109/SmartIndustryCon68821.2026.11492814

Uses

Direct Use

This corpus is intended for research and development in:

Training language models for the Tatar language, especially for low-resource scenarios
Fine-tuning multilingual models for Turkic languages
Text classification, named entity recognition, and other NLP tasks
Linguistic analysis of contemporary Tatar usage
Building applications for the Tatar-speaking community

Out-of-Scope Use

Redistribution of verbatim documents without permission from original copyright holders.
Any use that violates the rights of content creators whose texts are included.
Creation of misleading or harmful content based on the corpus.
Use as a source for factual or real-time information – the corpus is a static snapshot and may contain errors or outdated content.

Dataset Structure

The dataset is provided in two splits:

Split	Size	Description
full	2,465,867 documents	Complete corpus after deduplication
sample	1,000 documents	Small random sample for quick testing and prototyping

Field Descriptions

Field	Type	Description
`title`	string	Document title (available for 26.5% of documents)
`content`	string	Main text content
`category`	string	Topic category (available for 83.1% of documents)
`source`	string	Original source/website (available for 28.9% of documents)
`text`	string	Combined field: `title + "\n\n" + content`

Dataset Creation

Curation Rationale

Tatar is a low-resource Turkic language spoken by millions, yet it lacks large-scale text corpora necessary for modern NLP. This corpus was created to fill that gap, enabling the development of language technologies and supporting linguistic research for the Tatar-speaking community.

Source Data

Data Collection and Processing

Data were collected from publicly available Tatar-language websites, including:

News portals (Intertat, Tatar-inform, Azatliq)
Social media platforms (VK, Telegram)
Literary websites and digital libraries
Blogs and forums
Government and institutional websites

The processing pipeline involved:

Web crawling: Targeted crawling of known Tatar websites.
Extraction: HTML content was extracted and cleaned to plain text.
Deduplication: Soft deduplication with custom parameters:
- Strategy: custom
- Min group size: 800
- Keep percentage: 60%
- Prefix length: 300
- Result: from 3,032,454 input documents, 564,499 exact duplicates removed; large duplicate groups (1 group, 5,323 documents) processed, kept 3,235; final retention rate 81.3%.
Normalization: UTF-8 encoding, removal of extra whitespace.
Filtering: Empty and low-quality documents removed.

Who are the source data producers?

The original texts were produced by a wide range of authors and organisations: journalists, bloggers, social media users, writers, and government bodies. The dataset does not include personal identity information about the authors; it only contains the text itself.

Annotations

No additional annotations were added to the data. The fields category and source were extracted from the structure of the crawled websites (e.g., RSS feed categories, URL domains), not manually annotated.

Personal and Sensitive Information

The corpus consists of publicly available web content. While it may contain personal opinions or references to individuals, it does not deliberately collect private or sensitive data. Users should be aware that social media posts, in particular, may include personal information that was publicly shared.

Bias, Risks, and Limitations

Domain bias: A significant proportion of the corpus is health-related content (42.9% in the saqlyq category), which may skew language models trained on it.
Source coverage: Only 28.9% of documents have the source field; many documents lack metadata entirely.
Date information: No date metadata is available, so diachronic studies are not directly supported.
Web noise: Despite cleaning, some formatting artifacts and spam may remain.
Copyright constraints: The corpus contains copyrighted material; users must respect original authors' rights (see Licensing & Legal Notice).
Language style: Informal language, slang, and code-switching may be present, especially in social media content.

Recommendations

When training models, consider balancing the category distribution to mitigate domain bias.
Use the source field, where available, to filter or weight data based on provenance.
For any commercial application, verify the copyright status of the specific documents you intend to use.
Report any issues (e.g., low-quality texts, mis-categorized documents) via the Hugging Face repository.

Citation

If you use this dataset, please cite:

BibTeX (dataset):

@misc{tatar-web-corpus-2026,
  author = {Arabov, Mullosharaf Kurbonovich and TatarNLPWorld},
  title = {Tatar Web Corpus: A 251M Token Corpus for Low-Resource Turkic Language},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/TatarNLPWorld/tatar-web-corpus-v3}
}

APA (dataset): TatarNLPWorld. (2026). Tatar Web Corpus: A 251M Token Corpus for Low-Resource Turkic Language. Hugging Face. https://huggingface.co/datasets/TatarNLPWorld/tatar-web-corpus-v3

Related Publications

The following papers by the TatarNLPWorld team also used this corpus:

BibTeX:

@inproceedings{arabov2026combining,
  author = {Arabov, M. K. and Gilmullin, R. A. and Burnashev, R. A.},
  title = {Combining Classical and Transformer-based Approaches for Text Classification and Topic Modeling of the Tatar Language},
  booktitle = {2026 IEEE Ural-Siberian Conference on Biomedical Engineering, Radioelectronics and Information Technology (USBEREIT)},
  year = {2026},
  pages = {1--4},
  doi = {10.1109/USBEREIT70063.2026.11580632}
}

@inproceedings{burnashev2026parameter,
  author = {Burnashev, R. A. and Gilmullin, R. A. and Arabov, M. K.},
  title = {Parameter-efficient fine-tuning of large language models for Tatar language: a comparative study of LoRA and QLoRA},
  booktitle = {2026 IEEE Ural-Siberian Conference on Biomedical Engineering, Radioelectronics and Information Technology (USBEREIT)},
  year = {2026},
  pages = {1--4},
  doi = {10.1109/USBEREIT70063.2026.11580544}
}

@inproceedings{arabov2026subword,
  author = {Arabov, M. K. and Gainullin, R. N. and Khusaenov, A. I.},
  title = {Modelling Subword Embeddings of Low-Resource Languages for the Digitalisation of Industry 4.0 Industrial Systems},
  booktitle = {2026 International Russian Smart Industry Conference (SmartIndustryCon)},
  year = {2026},
  pages = {74--79},
  doi = {10.1109/SmartIndustryCon68821.2026.11492814}
}

APA:

Arabov, M. K., Gilmullin, R. A., & Burnashev, R. A. (2026). Combining Classical and Transformer-based Approaches for Text Classification and Topic Modeling of the Tatar Language. In 2026 IEEE Ural-Siberian Conference on Biomedical Engineering, Radioelectronics and Information Technology (USBEREIT) (pp. 1–4). IEEE. https://doi.org/10.1109/USBEREIT70063.2026.11580632
Burnashev, R. A., Gilmullin, R. A., & Arabov, M. K. (2026). Parameter-efficient fine-tuning of large language models for Tatar language: a comparative study of LoRA and QLoRA. In 2026 IEEE Ural-Siberian Conference on Biomedical Engineering, Radioelectronics and Information Technology (USBEREIT) (pp. 1–4). IEEE. https://doi.org/10.1109/USBEREIT70063.2026.11580544
Arabov, M. K., Gainullin, R. N., & Khusaenov, A. I. (2026). Modelling Subword Embeddings of Low-Resource Languages for the Digitalisation of Industry 4.0 Industrial Systems. In 2026 International Russian Smart Industry Conference (SmartIndustryCon) (pp. 74–79). IEEE. https://doi.org/10.1109/SmartIndustryCon68821.2026.11492814

Glossary

Token: An estimated unit of text based on word count (1 word ≈ 1.2 tokens).
Soft deduplication: A method that removes exact duplicates while sampling from large duplicate groups to preserve diversity.
Low-resource language: A language with limited digital resources and computational tools available.

More Information

Top 10 Categories

Category	Count	Percentage
saqlyq (health)	1,057,397	42.9%
ialgyz16	524,453	21.3%
sovet (advice)	139,582	5.7%
allahtan (from God)	79,174	3.2%
Poland visa points	33,301	1.4%
Uncategorized	27,624	1.1%
Mexico visa points	25,860	1.0%
Turkey visa points	24,726	1.0%
Italy visa points	17,997	0.7%
Society	8,963	0.4%

Top 10 Words

Word (Tatar)	English translation	Count
торак	housing, dwelling	884,971
урнашкан	located, situated	700,638
һәм	and	520,411
буенча	according to, along	501,740
төркем	group, category	414,984
әлифба	alphabet	352,870
инде	already, now	332,946
уртача	average, medium	329,648
пунктларытөркем	group of points / items group	316,992
да	also, too (particle)	286,006

Dataset Card Authors

Mullosharaf K. Arabov (TatarNLPWorld)
Contributions from the TatarNLPWorld community.

Dataset Card Contact

For questions, takedown requests, or collaboration, please open an issue on the Hugging Face repository:
https://huggingface.co/datasets/TatarNLPWorld/tatar-web-corpus-v3/discussions
📧 Direct email: marabov@kpfu.ru

Downloads last month: 13

Total file size:

2.42 GB