TajikNLPWorld/tajik-wiki-corpus

Dataset Card for Tajik Wikipedia Corpus

This dataset contains the complete dump of the Tajik Wikipedia (tgwiki) in Cyrillic script, processed and cleaned for natural language processing tasks. The corpus includes articles with their titles, categories, and metadata.

Dataset Details

Dataset Description

The Tajik Wikipedia Corpus is a collection of 79,985 Wikipedia articles in the Tajik language, totalling approximately 101.6 million characters and 15.3 million words. The data has been extracted from the Tajik Wikipedia dump and processed to ensure clean, well-structured text suitable for NLP applications.

Field	Description
title	Article title (present in 100% of documents)
text	Full article text (main content)
category	Wikipedia categories (comma-separated, present in 90.9% of articles)
source	Fixed as "tgwiki" for all documents
date	Publication date (not available in this version)
url	Original Wikipedia URL (present in 100% of documents)

The corpus is split into two subsets:

• full: Complete corpus with all 79,985 Wikipedia articles

• sample: Random 1,000‑article subset for quick testing and development

• Curated by: Arabov Mullosharaf Kurbonovich

• Language(s): Tajik (tg), Cyrillic script

• License: CC BY-SA 4.0 (compatible with Wikipedia's license)

Uses

Direct Use

The dataset is ideal for:

• Language modeling and pre-training Transformer models for Tajik
• Text classification (category prediction, topic modeling)
• Named entity recognition (Wikipedia contains rich named entities)
• Information retrieval and semantic search
• Cross-lingual learning (with Persian, Russian, or other languages)
• Lexical and linguistic studies of Tajik

Out-of-Scope Use

The dataset should not be used for:

• Temporal analysis (articles lack publication dates)
• Identifying individuals beyond public figures mentioned in Wikipedia
• Tasks requiring real-time updates (this is a static snapshot)

Dataset Structure

Data Fields

Field	Type	Coverage	Description
title	string	100%	Wikipedia article title
text	string	100%	Full article content (main text)
category	string	90.9%	Wikipedia categories (comma-separated)
source	string	100%	Always "tgwiki" for this dataset
date	string (YYYY-MM-DD)	0%	Not available in this version
url	string	100%	Original Wikipedia URL (e.g., https://tg.wikipedia.org/wiki/...)

Data Splits

Split	Documents	Size (approx)	Description
full	79,985	~100 MB	Complete Wikipedia corpus
sample	1,000	~1.3 MB	Random sample for testing/development

📊 Corpus Statistics

Overall Statistics

Metric	Value
Total documents	79,985
Total characters	101,554,357
Total words	15,327,781
Approx. tokens	18,327,348
Unique words (vocabulary)	599,402
Average document length	1,270 characters / 192 words
Average title length	17 characters / 2.4 words
Average text length	1,253 characters / 189 words

Field Coverage

Field	Count	Percentage
title	79,985	100.0%
category	72,731	90.9%
source	79,985	100.0%
date	0	0.0%
url	79,985	100.0%

Document Length Distribution

0-500 chars:      21,103 (26.4%) █████████████
500-1000 chars:   23,348 (29.2%) ██████████████
1000-2000 chars:  24,793 (31.0%) ███████████████
2000-5000 chars:   9,004 (11.3%) █████
5000-10000 chars:  1,301 (1.6%)  █
10000+ chars:        436 (0.5%)  

Top 20 Categories

#	Category	Count	Percentage
1	Фурудгоҳҳо аз рӯи алифбо (Airports by alphabet)	9,714	12.14%
2	Ниммуҳофизатгарони футбол (Football defenders)	7,650	9.56%
3	Муҳофизатгарони футбол (Football defenders)	6,058	7.57%
4	Маҳалҳои аҳолинишин аз рӯи алифбо (Populated places by alphabet)	5,307	6.63%
5	Ҳуҷумкунандагони футбол (Football forwards)	5,099	6.37%
6	Футболбозони Бразилия (Brazilian footballers)	4,178	5.22%
7	Футболбозони Испания (Spanish footballers)	3,859	4.82%
8	Деҳаҳои Тоҷикистон (Villages of Tajikistan)	3,821	4.78%
9	Фурудгоҳҳои Иёлоти Муттаҳидаи Амрико (US airports)	3,402	4.25%
10	Футболбозони Аргентина (Argentine footballers)	3,090	3.86%
11	Футболбозони Русия (Russian footballers)	2,558	3.20%
12	Ҳавогардҳо (Airlines)	2,544	3.18%
13	Мақолоти ҳавогардҳо (Airline articles)	2,354	2.94%
14	Мураббиёни футбол (Football coaches)	1,749	2.19%
15	Дарвозабонони футбол (Football goalkeepers)	1,653	2.07%
16	Фурудгоҳҳои Канада (Canadian airports)	1,422	1.78%
17	Маҳаллаҳои аҳолинишин аз рӯи алифбо (Localities by alphabet)	1,267	1.58%
18	Омӯзгорони Тоҷикистон (Tajikistan educators)	1,159	1.45%
19	Олимони Тоҷикистон (Tajikistan scientists)	1,087	1.36%
20	Қатъномаҳои Шӯрои Амният (UN Security Council resolutions)	1,049	1.31%

Category Distribution Insights

The category distribution reveals that the Tajik Wikipedia has a strong focus on:

• Football/sports content (~30% of articles)
• Geography (airports, populated places, villages)
• Biographies (footballers, educators, scientists)

This reflects the collaborative nature of Wikipedia where sports and geography articles are often systematically created.

Vocabulary Statistics

Metric	Value
Unique words	599,402
Type-token ratio	0.039
Most frequent words	(common Tajik function words and Wikipedia-specific terms)

Dataset Creation

Source Data

The corpus was created from the Tajik Wikipedia dump (tgwiki) in Cyrillic script. Wikipedia articles were extracted and processed to create a clean, structured dataset.

Processing Steps

1. Extraction: Articles were extracted from the Wikipedia XML dump
2. Content cleaning: HTML tags, templates, and formatting were removed
3. Text normalization: Unicode normalization and character cleaning
4. Category processing: Categories were extracted and cleaned
5. Deduplication: Duplicate articles were removed
6. Quality filtering: Articles with insufficient content were filtered out

Personal and Sensitive Information

Wikipedia articles may contain information about living persons. The dataset includes public information as it appears on Wikipedia. Users should be aware of Wikipedia's content policies and potential privacy considerations.

Bias, Risks, and Limitations

Known Biases

• Content bias: Strong representation of sports (especially football) and geography articles due to Wikipedia's content creation patterns
• Geographic bias: Tajikistan and Central Asian content is well-represented; global coverage is less balanced
• Gender bias: Like many Wikipedias, there may be gender imbalances in biographical articles
• Temporal bias: Articles reflect the state of Wikipedia at the time of extraction (no update history)

Missing Data

• Dates: No publication or last-updated dates are available
• Categories: 9.1% of articles lack categories (mostly stubs or new articles)
• Full revision history: Only the current version is included

Recommendations

• Use the category field for topic classification and filtering
• Consider the sports bias when training general-purpose models
• For historical analysis, use the full Wikipedia revision history (not included here)
• For gender-balanced studies, consider subsampling or using additional sources

Usage Examples

Loading the Dataset

from datasets import load_dataset

# Load the full corpus
dataset = load_dataset("TajikNLPWorld/tajik-wiki-corpus", split="full")

# Or load only the sample for testing
sample = load_dataset("TajikNLPWorld/tajik-wiki-corpus", split="sample")

# Access an article
article = dataset[0]
print(f"Title: {article['title']}")
print(f"Categories: {article['category']}")
print(f"URL: {article['url']}")
print(f"Text preview: {article['text'][:200]}...")

Filtering by Category

# Filter football-related articles
football_articles = dataset.filter(
    lambda x: 'футбол' in x['category'].lower()
)
print(f"Football articles: {len(football_articles)}")

Working with Categories

# Get all unique categories
all_categories = set()
for article in dataset:
    if article['category']:
        for cat in article['category'].split(','):
            all_categories.add(cat.strip())

print(f"Total unique categories: {len(all_categories)}")

Text Analysis Example

from collections import Counter

# Analyze word frequencies
word_counter = Counter()
for article in dataset.select(range(1000)):  # Analyze first 1000 articles
    words = article['text'].lower().split()
    word_counter.update(words)

print("Most common words:", word_counter.most_common(20))

Comparison with Full Web Corpus

Feature	Tajik Wikipedia Corpus	Tajik Web Corpus
Documents	79,985	319,298
Avg. length	1,270 chars	3,483 chars
Categories	90.9% coverage	50.9% coverage
Dates	0% coverage	0.3% coverage
Content style	Encyclopedia	News + mixed
Source	Single (Wikipedia)	Multiple (news, social, books)

Citation

BibTeX:

@dataset{arabov_tajik_wiki_corpus_2026,
  author = {Arabov, Mullosharaf Kurbonovich},
  title = {Tajik Wikipedia Corpus},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/TajikNLPWorld/tajik-wiki-corpus}
}

APA:

Arabov, M. K. (2026). Tajik Wikipedia Corpus [Data set]. Hugging Face. https://huggingface.co/datasets/TajikNLPWorld/tajik-wiki-corpus

Glossary

• Wikipedia dump: A snapshot of all Wikipedia articles at a specific point in time
• Category: Wikipedia's classification system for organizing articles
• Cyrillic script: The writing system used for Tajik in Tajikistan
• Type-token ratio: A measure of vocabulary diversity (unique words ÷ total words)

Acknowledgements

We thank:

• The Tajik Wikipedia community for their incredible work maintaining and expanding the encyclopedia
• The Wikimedia Foundation for making Wikipedia data freely available
• The Hugging Face team for providing the platform to share this dataset

Dataset Card Authors

Arabov Mullosharaf Kurbonovich

Dataset Card Contact

cool.araby@gmail.com