TatCorp-222M is the largest open corpus of Tatar language texts prepared for research in Natural Language Processing (NLP) and language modelling. The corpus contains 541,543 documents and approximately 222 million tokens. It is provided in Parquet/JSONL format and optimized for streaming processing and training large language models.

Curated by: Arabov Mullosharaf Kurbonovich (Arabovs AI Research Lab)
Language(s) (NLP): Tatar (tt)
License: other – see Licensing & Legal Notice below.

Licensing & Legal Notice

This dataset is private and access is restricted:

Repository: arabovs-ai-lab/TatCorp_222M (private on Hugging Face Hub)
Access: Available upon request for academic and non-commercial research after signing a Data Use Agreement (DUA).
Redistribution: Redistribution of source texts is prohibited without explicit permission from the rights holders.
This dataset follows the practice established by large web‑crawled corpora such as HPLT and OSCAR:
- Original source texts (web pages, articles, Wikipedia entries, social media posts) remain the property of their respective authors and publishers. They are not owned by the dataset curator and are not covered by any open license.
- The structured compilation, metadata, and any original annotations created during dataset preparation are released under a custom license (see contact for details).
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/arabovs-ai-lab/TatCorp_222M (private)

Uses

Direct Use

The TatCorp-222M corpus can be used for:

Language modelling – fine‑tuning or training large language models for Tatar from scratch
Text classification – multi‑class categorization by topic, genre, or source
Topic modelling – discovering latent topics and their temporal dynamics
Text generation – training models for news, article, or creative text generation
Information retrieval – building semantic search systems for Tatar
Linguistic research – analysing morphology, syntax, and lexicon of modern Tatar

Out-of-Scope Use

Redistribution of source texts without explicit permission from rights holders.
Commercial use without signing a Data Use Agreement.
Any use that violates the rights of original content creators.
Identifying individuals or extracting personal information from the corpus.

Dataset Structure

Data Fields

Field	Type	Description
`id`	string	Unique identifier
`url`	string	Original source URL (100% coverage)
`title`	string	Document title (99.92% coverage)
`content`	string	Main text content (99.99% coverage)
`category`	string	Topic category (99.94% coverage, 482 unique categories)
`date`	string	Publication date (16.7% coverage)
`lang`	string	Language code (`tt`)
`source`	string	Source domain (e.g., `wikipedia.org`)
`license`	string	Original license of the source (if applicable)

Data Splits

Split	Size	Percentage
`train`	530,712	~98%
`validation`	5,415	~1%
`test`	5,416	~1%

Dataset Creation

Curation Rationale

Tatar is a low-resource Turkic language with limited large-scale corpora. This corpus was assembled to provide a substantial text resource for training modern NLP models, with a focus on diversity across news, Wikipedia, and social media content.

Source Data

Data Collection and Processing

Data were collected from publicly available Tatar-language websites. The processing pipeline included:

Web crawling of news portals, Wikipedia, social media, and blogs.
HTML extraction and cleaning to plain text.
Language filtering to retain only Tatar Cyrillic texts.
Deduplication and normalization.
PII removal – automatic and selective manual removal of personal data.
Splitting into train/validation/test partitions.

Who are the source data producers?

The original texts were produced by journalists, Wikipedia contributors, bloggers, and social media users. The compilation and structuring were performed by Arabovs AI Research Lab, but the intellectual content remains with the original authors and publishers.

Annotations

No additional manual annotations were added. The category field was derived from source metadata where available.

Personal and Sensitive Information

Automatic and selective manual removal of personal data (PII) was applied. However, complete removal is not guaranteed; additional verification is recommended for sensitive use cases.

Bias, Risks, and Limitations

Source imbalance: Wikipedia dominates (84.2%), which may bias the corpus towards encyclopedic style.
Content currency: Most dated documents are from 2020–2025; older periods are underrepresented.
Length variance: The standard deviation of document length is large (3,301.9 tokens), with some very long documents.
PII removal: Not guaranteed to be complete; users should verify before sensitive applications.
Copyright constraints: The underlying texts may be protected; redistribution requires permission (see Licensing & Legal Notice).

Recommendations

For balanced language modeling, consider sampling from different sources.
Use streaming mode for large-scale processing to manage memory.
Verify PII removal if the corpus will be used in sensitive contexts.
Contact the curator to sign a DUA before accessing the data.

Citation

BibTeX (dataset):

@dataset{arabov2026tatcorp,
  author = {Arabov, Mullosharaf Kurbonovich},
  title = {TatCorp-222M: A Large-Scale Tatar Language Corpus},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/arabovs-ai-lab/TatCorp_222M}
}

APA (dataset): Arabov, M. K. (2026). TatCorp-222M: A Large-Scale Tatar Language Corpus [Data set]. Hugging Face. https://huggingface.co/datasets/arabovs-ai-lab/TatCorp_222M

Glossary

Token – a unit of text (word or subword) used in NLP models.
Low-resource language – a language with limited digital resources and tools.
PII – Personally Identifiable Information.

More Information

Related Resources

TatarTokenizers – Trained tokeniser models and BPE/WordPiece configurations:
https://huggingface.co/arabovs-ai-lab/TatarTokenizers
Tatar2Vec – Trained word and document embedding models:
https://huggingface.co/arabovs-ai-lab/Tatar2Vec

Dataset Card Authors

Arabov Mullosharaf Kurbonovich (Arabovs AI Research Lab)

Dataset Card Contact

For access requests, DUA, takedown requests, or collaboration, please contact:
📧 Direct email: cool.araby@gmail.com

Downloads last month: 5

Total file size:

972 MB