Ossetian-News-Corpus

A clean, unlabeled corpus of Ossetian news articles designed for unsupervised learning tasks: clustering, embedding training, language modeling, and semantic search.
Contains approximately 12,000 curated records (filtered from a raw pool of 59,375 articles) in the Ossetian language, with metadata including source URL, publication date, media outlet, and author.

🔗 This is the source corpus for the following supervised datasets:
• Ossetian-News-Binary — binary classification (Sport vs Other)
• Ossetian-News-Multiclass — 12-class classification
• Ossetian-News-Multilabel — multi-label classification

📊 Dataset Statistics

Core Metrics

Metric	Value
Curated records	12,000
Raw pool size	59,375
Language	Ossetian (ISO 639-1: os)
Task type	Unsupervised / feature extraction
License	CC BY 4.0

Text Properties (Character-Level)

Metric	Value
Avg. length	1,459 characters
Median length	795 characters
Length range	1 – 98,791 characters

Token Statistics (cl100k_base encoding — GPT-4 tokenizer)

Metric	Raw Pool (59,375)	Curated Subset (12,000)
Total tokens	~63.7M	~12.9M*
Avg. tokens per record	1,073	~1,073
Median tokens per record	~450*	~450*
Max tokens (single record)	~21,000*	~8,500*

Estimated based on raw pool distribution; exact values depend on filtering criteria

💡 Why such a big difference?
Simple "split by spaces" counting gives 12.3M tokens for the raw pool, but TikToken (cl100k_base) — used by GPT-4, Llama 3, and many modern models — produces **63.7M tokens** (5.19× more). This is because:
• Ossetian uses Cyrillic with special characters (Ӕ, Гъ, Къ, etc.) that are tokenized separately
• Punctuation, URLs, and mixed-language fragments increase token count
• Always use tokenizer-specific counts for budgeting inference/training costs

Temporal Coverage

Metric	Value
Date range	November 2015 – April 2026
Unique days with content	3,318
Records without date	208 (0.3%)

Source Distribution (Raw Pool: 59,375)

Source	Records	%
iryston.tv	31,988	53.9%
xurzarin.ru	8,930	15.0%
sputnik-ossetia.com	8,715	14.7%
rastdzinad.ru	7,636	12.9%
alaniatv.ru	1,898	3.2%
mahdug.ru	208	0.4%

Author Diversity (Raw Pool)

Metric	Value
Unique named authors	29
Records with author	~67%
Records without author	~33% (19,639)
Top authors	Dzebisón Gulo (11.2%), Diana Kessaeva (5.0%), Zarina Batsazova (1.5%)

📁 Data Fields

Field	Type	Description
text	string	Full news text (headline + body content) in Ossetian
url	string	Original source URL of the news article
date	string	Publication date (ISO format: YYYY-MM-DD, when available)
site	string	Source media outlet domain
author*	string	Author name (when available; may be missing or generic)

*Included in raw pool; filtered in curated subsets

💻 Quick Start

from datasets import load_dataset

# Load and preview 5 samples
ds = load_dataset("OssetianNLPWorld/ossetian-news-corpus")
for i, sample in enumerate(ds["train"].select(range(5)), 1):
    print(f"\n--- Sample {i} ---")
    print(f"Site: {sample['site']} | Date: {sample.get('date', 'N/A')}")
    print(f"Text: {sample['text'][:200]}...")
    print(f"URL: {sample['url']}")

🔗 Related Datasets (Derived from this Corpus)

These supervised datasets were created by filtering and labeling subsets of this corpus:

Dataset	Task	Records	Link
Ossetian-News-Binary	Binary classification	3,296	🔗
Ossetian-News-Multiclass	12-class classification	13,036	🔗
Ossetian-News-Multilabel	Multi-label classification	13,036	🔗

🔗 Data Sources

News articles were collected from the following Ossetian-language media outlets:

Source	Type
alaniatv.ru	TV & news portal
iryston.tv	Regional TV & news
mahdug.ru	Cultural & social news
rastdzinad.ru	Regional news
sputnik-ossetia.com	News agency
xurzarin.ru	Independent news portal

🙏 About the Creator

This dataset and the OssetianNLPWorld organization on Hugging Face were created and are maintained by:

Arabov Mullosharaf Kurbonovich
Candidate of Physical and Mathematical Sciences, Associate Professor
Department of Data Analysis and Programming Technologies
Kazan Federal University, Russia

The OssetianNLP initiative is a personal research project aimed at developing natural language processing tools and resources for the Ossetian language and other low-resource languages of the Caucasus region.

📄 Citation

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

@dataset{arabov2026ossetiannewscorpus,
  author = {Arabov, Mullosharaf Kurbonovich},
  title = {{Ossetian-News-Corpus}: An Unlabeled Corpus of Ossetian News Articles},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/OssetianNLPWorld/ossetian-news-corpus},
  license = {CC BY 4.0}
}

⚠️ Limitations & Considerations

• Source imbalance: ~54% of raw data comes from iryston.tv; consider stratified sampling for balanced experiments.
• Text length variance: Extreme outliers (up to 98K characters); apply truncation or chunking for transformer models.
• Temporal gaps: Irregular publication frequency; not suitable for strict time-series without preprocessing.
• Author metadata: ~33% of records lack author information; use author field with caution.
• Language mixing: Some articles contain Russian code-switching or boilerplate metadata.
• Domain specificity: Content is news-focused; embeddings may not transfer well to social media or literary texts.

📬 Contact

For questions, collaboration proposals, or feedback:

Arabov Mullosharaf Kurbonovich
📧 cool.araby@gmail.com
🔗 Telegram: @cool_araby • VK: @cool.araby
🏢 Kazan Federal University, Kazan, Russia

---

Dataset version: 1.0 • Last updated: April 2026 • Part of the OssetianNLP initiative by Arabov M.K.