The Ossetian Web Corpus is a comprehensive collection of Ossetian-language texts gathered from three main sources: online news portals, Wikipedia articles, and digitized books. The corpus is designed to support natural language processing (NLP) research and development for the Ossetian language, a low-resource language spoken in the Caucasus region.

Curated by: OssetianNLPWorld
Language(s) (NLP): Ossetian (ISO 639-1: os)
License: Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0)

Dataset Sources

The data originates from three distinct domains:

News Articles – scraped from the Ossetian-language news site alaniatv.ru and similar regional news portals. These texts cover a wide range of contemporary topics.
Wikipedia – articles extracted from the Ossetian Wikipedia (Cyrillic edition). The articles span encyclopedic knowledge in Ossetian.
Books – digitized Ossetian books (fiction and non-fiction) sourced from a local virtual library of Ossetian literature.

All texts were processed to remove boilerplate content, publisher information, tables of contents, and non‑linguistic artifacts.

Repository: https://huggingface.co/datasets/OssetianNLPWorld/ossetian-web-corpus

Uses

Direct Use

This corpus is intended for:

Training or fine‑tuning language models for Ossetian.
Studying Ossetian morphology, syntax, and lexical distribution.
Building text classification, named entity recognition, or summarization systems for Ossetian.
Cross‑lingual transfer learning with other languages that use Cyrillic script.

Out-of-Scope Use

The corpus contains raw, unannotated text. It is not suitable for tasks requiring token‑level or sentence‑level annotations (e.g., part‑of‑speech tagging, parsing) without additional labeling. Because it includes news articles, it may contain time‑sensitive or politically charged content that should be handled with care in certain applications.

Dataset Structure

The dataset is provided in JSON Lines format with the following fields:

Field	Type	Description
`title`	string	Title of the document (news headline, Wikipedia article title, book chapter title)
`text`	string	Full text of the document. For convenience, this field contains the concatenation of `title` and the original body text, separated by a double newline (`\n\n`)
`category`	string	Thematic category (e.g., "Politics", "Culture", "History") – may be empty for some sources
`source`	string	Origin of the document: a news domain (e.g., `alaniatv.ru`), `"Википедия"` (Wikipedia), or `"Книга"` (Book)
`date`	string	Publication date in ISO 8601 format (e.g., `2019-01-25T15:00:00`). Empty for Wikipedia entries
`url`	string	Original URL or file path of the source

Splits

full: The complete corpus containing 76,599 documents (~231 MB uncompressed).
sample: A random subset of 1,000 documents (~3 MB) for quick exploration and prototyping.

Dataset Creation

Curation Rationale

Ossetian is a low‑resource language with limited digital corpora available for NLP research. The goal of this project was to assemble a large, diverse, and clean text collection that can serve as a foundation for developing language technologies for Ossetian speakers.

Source Data

Data Collection and Processing

News: Articles were scraped using custom crawlers, preserving metadata (title, date, category, URL).
Wikipedia: A dump of the Ossetian Wikipedia (Cyrillic) was parsed, extracting article text and categories.
Books: Selected books in .fb2, .doc, and plain text formats were converted to UTF‑8 text.

Processing steps:

Language filtering – Documents were kept only if they contained Ossetian‑specific Cyrillic letters (ӕ/Ӕ) in the first 1,000 characters.
Deduplication – Duplicate documents were removed based on the SHA‑256 hash of the first 1,000 characters of the content.
Text cleaning – Removed carriage returns, tab characters, empty lines, table‑of‑contents lines (e.g., XX\t35), publisher information, and other non‑textual artifacts.
Empty document removal – Entries with neither a title nor a body were discarded.

The final corpus was exported as a JSONL file and then uploaded to the Hugging Face Hub.

Who are the source data producers?

News articles: journalists and editors of Ossetian online media.
Wikipedia: volunteer contributors to the Ossetian Wikipedia.
Books: Ossetian authors and publishers.

Annotations

This dataset does not contain additional annotations. It is a raw text corpus.

Personal and Sensitive Information

The corpus consists of publicly available news articles, encyclopedia entries, and published books. No personal user data was collected. The texts may occasionally contain names of public figures or references to historical events, but no systematic anonymization was performed.

Bias, Risks, and Limitations

Domain bias: The corpus is skewed toward news and encyclopedic content; other registers (e.g., informal conversation, social media) are underrepresented.
Temporal bias: News articles cover a specific time range (mostly 2019–2025), which may affect language model behavior when applied to historical or future texts.
Geographic and cultural bias: The content primarily reflects the North Ossetian dialect and regional topics. South Ossetian and diaspora varieties are less represented.
Quality: While extensive cleaning was performed, some residual noise (e.g., OCR errors in books) may remain.

Recommendations

Users should be aware of the biases above and consider them when deploying models trained on this corpus. For applications requiring a balanced representation, additional data collection or filtering may be necessary.

Citation

If you use this dataset in your research, please cite it as:

BibTeX:

@dataset{ossetian_web_corpus,
  author       = {Arabov Mullosharaf Kurbonovich and OssetianNLPWorld},
  title        = {Ossetian Web Corpus},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/OssetianNLPWorld/ossetian-web-corpus}},
  note         = {Version 1.0}
}

APA: OssetianNLPWorld. (2025). Ossetian Web Corpus (Version 1.0) [Dataset]. Hugging Face. https://huggingface.co/datasets/OssetianNLPWorld/ossetian-web-corpus

Glossary

Ossetian: An Eastern Iranian language spoken in the Caucasus, written in an extended Cyrillic alphabet that includes the letter ӕ.
Token: The atomic unit used by tokenizers such as XLM‑RoBERTa.
Based on the xlm-roberta-base tokenizer:
- Text only (main content): 60,289,396 tokens (average 787.1 tokens per document).
- Title + text combined: 61,983,750 tokens (average 809.2 tokens per document).
Document length distribution (title + text, in tokens):
- 0: 8,533 documents (11.1%) – effectively empty records
- 1–128: 10,105 documents (13.2%)
- 129–256: 22,814 documents (29.8%)
- 257–512: 21,287 documents (27.8%)
- 513–1024: 8,834 documents (11.5%)
- 1025–2048: 3,962 documents (5.2%)
- 2049–4096: 925 documents (1.2%)
- 4097–8192: 139 documents (0.2%)
- >8192: 0 documents (0.0%)

Dataset Card Authors

OssetianNLPWorld

Dataset Card Contact

For questions or collaboration, please open an issue on the dataset repository.

Downloads last month: 20

Total file size:

107 MB