Dataset Card for Ossetian-Wikipedia

Quick Summary

A cleaned, structured corpus of 17,889 articles from the Ossetian Wikipedia (os.wikipedia.org), stripped of MediaWiki markup and enriched with linguistic metadata: script detection, category lists, word/sentence counts, and readability scores. Intended for low-resource language modelling, tokenizer research, and script/language detection tasks.

Dataset Details

Dataset Description

Ossetian (ISO 639-1: os, ISO 639-3: oss, endonym ирон æвзаг) is an Iranian language spoken by roughly 500,000 people, primarily in North and South Ossetia. It uses a Cyrillic-based alphabet containing the unique letter ӕ (U+04D5, Cyrillic small letter schwa), which most mainstream tokenizers do not handle efficiently.

This dataset provides the full text corpus of the Ossetian Wikipedia as of the latest Wikimedia dump, prepared for downstream NLP research. Each record contains the article title, cleaned plain text, a category list, and automatically derived script and quality metadata.

• Curated by: Arabov Mullosharaf Kurbonovich (OssetianNLPWorld)
• Funded by [optional]: Independent research — no external funding
• Shared by [optional]: OssetianNLPWorld
• Language(s) (NLP): Ossetian (os)
• License: CC BY-SA 4.0 (inherited from the Wikimedia source)

Dataset Sources [optional]

• Repository: https://huggingface.co/datasets/OssetianNLPWorld/ossetian-wikipedia
• Paper [optional]: N/A
• Demo [optional]: N/A
• Source dump: https://dumps.wikimedia.org/oswiki/latest/

Uses

Direct Use

• Continual pre-training or supervised fine-tuning of low-resource language models (encoder or decoder).
• Training or evaluating tokenizers (BPE, Unigram, WordPiece) for Cyrillic languages — see the token statistics below.
• Script detection and language identification benchmarks.
• Category classification, text-quality estimation, readability analysis.
• Benchmarking tokenization efficiency (chars/token) for under-represented scripts.

Out-of-Scope Use

• Pre-training a general-purpose LLM from scratch. 8.7M tokens is insufficient — the corpus supports fine-tuning / continual pre-training, not from-scratch training.
• Instruction tuning. No user/assistant pairs are included.
• Factual QA over Wikipedia content. Text may be outdated or contain wiki-specific artefacts.
• PII extraction or de-anonymization. Articles mention public figures; do not use the corpus to build person profiles.
• Machine translation without a parallel corpus. This dataset is monolingual.

Dataset Structure

Fields

Field	Type	Description
title	string	Article title (Wikipedia canonical form)
text	string	Plain text after MediaWiki markup removal
categories	List[string]	Categories declared in the article
alphabet	string	Detected script: cyrillic, latin, mixed, other
categories_count	int32	Number of categories
text_length	int32	Character count of text
url	string	Canonical URL of the article
alphabet_confidence	float32	Confidence of script detection (0–1)
cyr_ratio	float32	Fraction of Cyrillic letters in text
lat_ratio	float32	Fraction of Latin letters in text
word_count	int32	Word count in text
sentence_count	int32	Sentence count
readability_score	float32	Heuristic readability (0–100)
avg_sentence_length	float32	Average sentence length (words)

Splits

Single split: train (all 17,889 articles). No validation/test split is provided — create your own (e.g. 90/5/5), ideally holding out by url to ensure exact reproducibility.

Alphabet Distribution

The corpus is overwhelmingly Cyrillic (Ossetian’s native script). Articles whose cleaned text is dominated by Latin letters — usually proper names, code blocks, or reference-only stubs — are re-labelled latin; mixed-script articles become mixed. Per-alphabet JSONL files ship under raw/:

• raw/oswiki_cyrillic.jsonl — Cyrillic articles
• raw/oswiki_latin.jsonl — Latin articles
• raw/oswiki_mixed.jsonl — Mixed-script articles
• raw/oswiki_all_articles.jsonl — full corpus

Exact per-alphabet counts are in stats/detailed_stats.json.

Token Statistics

Measured on the full corpus (title + text) using tiktoken and transformers:

Tokenizer	Vocab	Title tokens	Text tokens	Combined	chars/token	vs English
GPT-4o (o200k_base)	200k	117,056	8,619,655	8,736,711	1.96	2.0×
Qwen2.5	151k	133,717	9,742,915	9,876,632	1.74	2.3×
Mistral-Nemo	131k	121,810	9,567,692	9,689,502	1.77	2.3×
GPT-4 (cl100k_base)	100k	152,175	10,568,731	10,720,906	1.60	2.5×
SmolLM2	49k	208,844	14,054,815	14,263,659	1.20	3.3×
GPT-2 (r50k_base)	50k	258,415	17,179,538	17,437,953	0.99	4.1×

Context fit (GPT-4o, combined):

max_length	Articles fitting without truncation
2,048	17,727 (99.1 %)
4,096	17,852 (99.8 %)
8,192	17,879 (99.9 %)
32,768	17,889 (100 %)

Distribution (GPT-4o, combined): mean = 488, p50 = 538, p95 = 889, p99 = 1,989, max = 22,264.

Why is chars/token so low? The letter ӕ (U+04D5), unique to Ossetian, is not a single token in most mainstream tokenizers and gets split into 2–3 byte-level tokens. This produces a ≈2× cost factor versus English. Adding ӕ/Ӕ as special tokens reduces the token count by 10–15 % (see Recommendations).

Dataset Creation

Curation Rationale

Ossetian is a low-resource Iranian language with a small but growing digital footprint. The Ossetian Wikipedia is the single largest openly-licensed curated text corpus in the language. Shipping a clean, metadata-enriched, ready-to-load version lowers the barrier for researchers who would otherwise have to parse a multi-gigabyte XML dump themselves.

Source Data

Data Collection and Processing

• Source: official Wikimedia dump oswiki-latest-pages-articles.xml.bz2 from dumps.wikimedia.org.
• Filtering:
  • Namespace 0 (main articles) only.
  • Redirects (#REDIRECT, #перенаправление) removed.
  • Articles whose cleaned text is shorter than 100 characters dropped.
• Cleaning:
  • HTML comments, {{...}} templates, <ref> tags, and other HTML removed.
  • Internal wiki links converted to display text; images and external links stripped.
  • Bold/italic markup, headings, tables, and HTML entities removed.
  • Consecutive blank lines collapsed.
• Metadata (automatically computed):
  • Script detection via the Cyrillic/Latin letter ratio.
  • Category extraction via regex on [[Категори:...]] / [[Category:...]].
  • Word/sentence/readability metrics computed on cleaned text.
• Tools: Python 3, xml.etree.ElementTree (iterparse), bz2, requests, datasets, huggingface_hub, tiktoken, transformers.

Who are the source data producers?

The original text was produced collaboratively by volunteer editors of the Ossetian Wikipedia and is distributed under CC BY-SA 4.0. Cleaning, metadata generation, packaging, and hosting were performed by OssetianNLPWorld (Arabov Mullosharaf Kurbonovich).

Annotations [optional]

No manual annotations. All metadata (alphabet, categories, word_count, readability_score, etc.) is computed automatically during parsing.

Personal and Sensitive Information

The dataset inherits any PII present in the source Wikipedia — biographical entries, names of public figures, historical references. No additional anonymization was performed beyond what Wikipedia itself does. If you intend to use this corpus for person-centric tasks, apply your own PII scrubbing and comply with applicable data-protection laws (GDPR and similar).

Bias, Risks, and Limitations

• Wikipedia bias. Topical coverage reflects the interests and demographics of Wikipedia editors — Western-leaning, uneven depth, gaps in topics of local Ossetian interest.
• Historical and political bias. Certain entries (especially on Ossetia, Russia, and Georgia) may reflect contested narratives. Do not treat as ground truth.
• Cleaning artefacts. Regex-based cleaning removes templates aggressively — infoboxes, references, and some structured content are lost, and a small amount of residual markup may remain.
• Heuristic script detection. Short or code-heavy articles may be misclassified as latin or mixed.
• Small size. 17,889 articles / 8.7M tokens is small — sufficient for fine-tuning, insufficient for pre-training.
• Duplicate topics. Some articles may cover the same entities under different titles (surviving redirects).
• No train/val/test split. Create your own to avoid leakage.
• Language mixing. Ossetian Wikipedia includes Russian and English fragments (mostly names and technical terms) that may leak into monolingual training.

Recommendations

• Choose the tokenizer carefully. For training on this corpus, prefer Qwen2.5, Mistral-Nemo, or a Llama-3-family tokenizer. Avoid GPT-2 (2× token inflation).
• Add ӕ/Ӕ as special tokens when you control the tokenizer — this alone reduces the token count by 10–15 %.
• Use max_length = 2048 — it covers >99 % of articles without truncation.
• Enable packing for causal-LM training. Median article length is only ~538 tokens, so 3–4 short articles fit into one 2048-token window, doubling throughput versus padding.
• Hold out by url when creating custom splits for reproducibility.
• Deduplicate if you combine with related-language Wikipedia content (Persian, Tajik, Kurdish).

Citation [optional]

If you use this dataset in academic work, please cite:

BibTeX (dataset):

@dataset{arabov2026ossetianwikipedia,
  author = {Arabov, Mullosharaf Kurbonovich},
  title = {{Ossetian-Wikipedia}: An Open Corpus of Ossetian Wikipedia Articles},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/OssetianNLPWorld/ossetian-wikipedia}
}

APA (dataset): Arabov, M. K. (2026). Ossetian-Wikipedia: An Open Corpus of Ossetian Wikipedia Articles [Data set]. Hugging Face. https://huggingface.co/datasets/OssetianNLPWorld/ossetian-wikipedia

Glossary [optional]

• Ossetian – an Iranian language (ISO 639-1: os) spoken in the Caucasus, written in a modified Cyrillic alphabet.
• ӕ (schwa) – the unique Cyrillic letter U+04D5, present in Ossetian but not in Russian.
• Low-resource language – a language with limited digital text, models, and tools.
• chars/token – average number of source characters encoded per token; higher is more efficient.
• Mixed-script article – an article whose text contains a comparable amount of Latin and Cyrillic characters.
• Readability score – a heuristic 0–100 value derived from average sentence length; not a peer-reviewed metric.
• Wiki dump – the official XML snapshot of a Wikipedia edition published by the Wikimedia Foundation.

More Information

• Tokenisation statistics were computed with tiktoken and transformers across six tokenizers of varying vocabulary size. Scripts and additional analysis are available on request.
• For questions, contributions, or bug reports, please open an issue on the Hugging Face repository.

Dataset Card Authors

Arabov Mullosharaf Kurbonovich (OssetianNLPWorld)

Dataset Card Contact

For questions, takedown requests, or collaboration, please open an issue on the Hugging Face repository: https://huggingface.co/datasets/OssetianNLPWorld/ossetian-wikipedia/discussions

📧 Direct email: cool.araby@gmail.com