Dataset Card for Tajik Web Corpus
This dataset contains a large‑scale web corpus of the Tajik language, collected from various online sources (news portals, Wikipedia, social media, etc.). The corpus is intended for natural language processing tasks such as language modeling, text classification, and machine translation.
Dataset Details
Dataset Description
The Tajik Web Corpus is a collection of 319,298 documents in the Tajik language, totalling approximately 1.11 billion characters and 168.5 million words. The data has been cleaned, normalized, and deduplicated, and is provided in JSONL format with the following fields:
| Field | Description |
|---|---|
title |
Document title (present in 99.7% of documents) |
text |
Full document text (title + body when title exists) |
category |
Thematic category (present in 50.9% of documents) |
source |
Original source file name (100% coverage) |
date |
Publication date in YYYY-MM-DD format (present in 0.3% of documents) |
url |
Original URL (100% coverage) |
Note: The title field is kept separate from text in the dataset structure, though in the current version text already includes the title for convenience.
The corpus is split into two subsets:
full: Complete corpus with all 319,298 documentssample: Random 1,000‑document subset for quick testing and development
Dataset Sources
The corpus was compiled from publicly available sources, including:
- Tajik Wikipedia (
tgwiki_cyrillic) — encyclopedic articles - News portals: Asia-Plus (
asiaplus), Khovar (khovar), Ozodi (ozodiarticles), Farazh (farazh_news_dataset), Sadoi Mardum (sadoimardum), Jumhuriyat (jumhuriyat) - Government sources: Ministry of Foreign Affairs (
mfa-tj), official portals (dushanbe-tj,mit-tj,omuzgor) - Social media: Telegram posts (
khiradtj_posts) - Books and literature: Tajik books (
tajik_books_final), literary works
For a full list of sources, see the dataset statistics below.
- Curated by: Arabov Mullosharaf Kurbonovich
- Language(s): Tajik (tg), with some Persian (fa) content
- License: CC BY-SA 4.0
Uses
Direct Use
The dataset is intended for training and evaluating NLP models for the Tajik language, including:
- Language modeling (e.g., for pre‑training Transformer models like BERT, GPT)
- Text classification (e.g., topic classification, genre detection, sentiment analysis)
- Machine translation (Tajik ↔ Persian, Tajik ↔ Russian)
- Information retrieval and text mining
- Linguistic studies of the Tajik language (vocabulary, syntax, style)
Researchers may also use the corpus for comparative studies across Persian varieties.
Out-of-Scope Use
The dataset should not be used for:
- Identifying individuals or personal information (the data has been anonymized)
- Generating harmful or misleading content
- Tasks requiring fine‑grained temporal information (only 0.3% of documents have dates)
- Geographic analysis (content predominantly from Tajikistan)
Dataset Structure
Data Fields
| Field | Type | Coverage | Description |
|---|---|---|---|
title |
string | 99.7% | Document title (separate field for flexibility) |
text |
string | 100% | Full document text (includes title when available) |
category |
string | 50.9% | Thematic category (e.g., Сиёсат, Иқтисод, Ҷамъият) |
source |
string | 100% | Original source file name |
date |
string (YYYY-MM-DD) | 0.3% | Publication date (mainly from books corpus) |
url |
string | 100% | Original URL (where available) |
Data Splits
| Split | Documents | Size (JSONL) | Description |
|---|---|---|---|
full |
319,298 | 1.93 GB | Complete corpus |
sample |
1,000 | 6.1 MB | Random sample for testing/development |
📊 Corpus Statistics
Overall Statistics
| Metric | Value |
|---|---|
| Total documents | 319,298 |
| Total characters | 1,112,145,760 |
| Total words | 168,483,674 |
| Approx. tokens | 201,921,066 |
| Unique words (vocabulary) | 681,666 |
| Average document length | 3,483 characters / 528 words |
| Average title length | 55 characters / 8 words |
| Average text length | 3,428 characters / 520 words |
Field Coverage
| Field | Count | Percentage |
|---|---|---|
title |
318,187 | 99.7% |
category |
162,442 | 50.9% |
source |
319,298 | 100.0% |
date |
976 | 0.3% |
url |
319,298 | 100.0% |
Top 20 Categories
| Category | Count | Percentage |
|---|---|---|
| Ҷамъият (Society) | 26,006 | 8.14% |
| Иқтисод (Economy) | 14,249 | 4.46% |
| Сиёсат (Politics) | 13,058 | 4.09% |
| Ҳодиса (Events) | 9,607 | 3.01% |
| Амният (Security) | 8,746 | 2.74% |
| Хабарҳо (News) | 8,491 | 2.66% |
| Тоҷикистон (Tajikistan) | 6,515 | 2.04% |
| Фарҳанг (Culture) | 6,229 | 1.95% |
| Фурудгоҳҳо аз рӯи алифбо (Airports by alphabet) | 4,775 | 1.50% |
| Варзиш (Sports) | 4,430 | 1.39% |
| Ҳукумат (Government) | 2,810 | 0.88% |
| Ҳуқуқ (Law) | 2,746 | 0.86% |
| Ҷаҳон (World) | 2,737 | 0.86% |
| Маориф (Education) | 1,654 | 0.52% |
| Осиёи Марказӣ (Central Asia) | 1,608 | 0.50% |
| Минтақа (Region) | 909 | 0.28% |
| Русия (Russia) | 881 | 0.28% |
| Мақомоти давлатӣ (State authorities) | 868 | 0.27% |
| Фурудгоҳ (Airport) | 814 | 0.25% |
| Дигар (Other) | 31,927 | 10.00% |
Note: "Дигар" (Other) includes categories with less than 800 documents and documents without categories.
Top 20 Sources
| Source | Count | Percentage |
|---|---|---|
| asiaplus | 88,925 | 27.85% |
| tgwiki | 79,985 | 25.05% |
| khovar | 52,619 | 16.48% |
| sadoimardum | 21,879 | 6.85% |
| oila | 19,440 | 6.09% |
| ovozitojik | 17,143 | 5.37% |
| jumhuriyat | 8,921 | 2.79% |
| ozodi | 8,549 | 2.68% |
| mfa-tj | 7,783 | 2.44% |
| farazh | 7,133 | 2.23% |
| khiradtj | 4,048 | 1.27% |
| sssr | 1,754 | 0.55% |
| dushanbe-tj | 29 | 0.01% |
| mit-tj | 8 | 0.00% |
| sputnik | 74 | 0.02% |
| omuzgor | 32 | 0.01% |
| books_corpora | 976 | 0.31% |
Text Length Distribution
Document length distribution:
0-500 chars: 12,847 (4.0%)
500-1000 chars: 24,357 (7.6%)
1000-2000 chars: 51,943 (16.3%)
2000-5000 chars: 119,472 (37.4%)
5000-10000 chars: 72,314 (22.7%)
10000+ chars: 38,365 (12.0%)
Average: 3,483 chars | Median: 2,847 chars
Dataset Creation
Curation Rationale
The corpus was created to address the lack of large‑scale, clean, and publicly available text data for the Tajik language. By aggregating diverse sources and applying extensive cleaning and normalization, the dataset provides a solid foundation for NLP research in Tajik.
Data Collection and Processing Pipeline
Data was collected from publicly accessible websites and archives. The following steps were applied:
- Scraping / Collection: Documents were collected from various sources (Wikipedia, news sites, social media, etc.)
- Initial Cleaning: HTML tags, URLs, email addresses, and non‑textual elements were removed
- Character Normalization: Non‑standard characters were mapped to their correct Tajik equivalents
- Quality Filtering: Documents with fewer than 20 characters or predominantly non‑Tajik content were filtered out
- Title Cleaning: Garbage titles (numbers only, URLs, hashes, etc.) were removed
- Category Normalization: Categories were cleaned and unified where possible
- Final Assembly: All processed documents were merged into a single JSONL file
Who are the source data producers?
The source data was produced by a variety of entities:
- Wikipedia: community‑contributed articles
- News portals: professional journalists and media outlets (Asia-Plus, Khovar, Ozodi, etc.)
- Government sites: press releases and official documents
- Social media: public posts (Telegram, etc.)
- Books and literature: digitized works
Personal and Sensitive Information
Efforts were made to exclude personally identifiable information. However, the corpus contains news articles that may mention names of public figures. Users should be aware that the dataset may include such references.
Bias, Risks, and Limitations
Known Biases
- Genre bias: The corpus is dominated by news content (
80%) and Wikipedia (25%), which may bias models towards journalistic style. Social media and literary texts are underrepresented - Geographic bias: Most sources originate in Tajikistan; content from other Tajik‑speaking regions (e.g., Afghanistan) is limited
- Temporal bias: The data spans several years but lacks comprehensive dating, making time‑based analysis unreliable
- Category bias: Some categories (e.g., "Ҷамъият", "Иқтисод") are overrepresented while others are sparse
Missing Data
- Categories: Available for only 50.9% of documents
- Dates: Available for only 0.3% of documents (mainly from the books corpus)
- Titles: 0.3% of documents lack titles (mostly from khiradtj and tgwiki sources)
Recommendations
Users should consider these limitations when using the dataset:
- For genre-balanced training, consider subsampling or combining with other corpora
- When using categories, be aware that 49.1% of documents have empty category fields
- For time‑based analysis, the dataset is not suitable due to sparse date information
- For title-based tasks, note that 99.7% of documents have titles available
Usage Examples
Loading the Dataset
from datasets import load_dataset
# Load the full corpus
dataset = load_dataset("TajikNLPWorld/tajik-web-corpus", split="full")
# Or load only the sample for testing
sample = load_dataset("TajikNLPWorld/tajik-web-corpus", split="sample")
# Access a document
doc = dataset[0]
print(f"Title: {doc['title']}")
print(f"Category: {doc['category']}")
print(f"Source: {doc['source']}")
print(f"Text preview: {doc['text'][:200]}...")
Filtering by Category
# Filter documents by category
politics_articles = dataset.filter(lambda x: x['category'] == 'Сиёсат')
print(f"Politics articles: {len(politics_articles)}")
Working with Splits
# Load both splits
dataset = load_dataset("TajikNLPWorld/tajik-web-corpus")
full = dataset["full"]
sample = dataset["sample"]
print(f"Full size: {len(full):,}")
print(f"Sample size: {len(sample):,}")
Citation
BibTeX:
@dataset{arabov_tajik_web_corpus_2026,
author = {Arabov, Mullosharaf Kurbonovich},
title = {Tajik Web Corpus},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/TajikNLPWorld/tajik-web-corpus}
}
APA:
Arabov, M. K. (2026). Tajik Web Corpus [Data set]. Hugging Face. https://huggingface.co/datasets/TajikNLPWorld/tajik-web-corpus
Glossary
- Tajik language: A variety of Persian spoken primarily in Tajikistan, written in the Cyrillic script
- Category: Thematic label assigned to the document (e.g., "Politics", "Economy")
- Source: The original file name from which the document was taken
- Full split: Complete dataset with all documents
- Sample split: Random subset for testing (1,000 documents)
Acknowledgements
We thank:
- The Tajik Wikipedia community for their valuable contributions
- News portals (Asia-Plus, Khovar, Ozodi, Farazh, etc.) for making their content publicly available
- The Hugging Face team for providing the platform to share this dataset
- All contributors who helped with data collection and cleaning
Dataset Card Authors
Arabov Mullosharaf Kurbonovich
Dataset Card Contact
- Downloads last month
- 27