This dataset contains 109,082 Tajik news articles collected from 8 online sources. It is designed for clustering, embedding learning, and unsupervised NLP tasks. Each record includes the full text, title, normalized category, source, length, URL, and publication date.

The articles have been cleaned, deduplicated, and filtered to ensure a minimum and maximum content length (50–10,000 characters). The original categories have been normalized into 54 coherent labels (e.g., Ҷамъият, Иқтисод, Сиёсат). Empty or uninformative categories have been excluded.

Dataset Sources

asiaplus.tj – 67,813 articles
ovozitojik.tj – 16,636 articles
mfa.tj – 7,584 articles
ozodi.org – 6,626 articles
farazh.tj – 5,411 articles
jumhuriyat.tj – 3,716 articles
sssr.tj – 1,222 articles
sputnik.tj – 74 articles

All data were collected between 2015 and 2025 (exact date ranges vary by source).

Uses

Direct Use

The dataset can be used for:

Clustering news articles by topic
Learning document embeddings (e.g., with Sentence-BERT, LASER)
Evaluating unsupervised topic modeling methods
Pre‑training or fine‑tuning language models for Tajik

Out-of-Scope Use

The dataset is not intended for supervised classification tasks without further splitting (a separate dataset for classification is available as TajikNLPWorld/tajik-news-multiclass). It also does not contain parallel translations or audio.

Dataset Structure

Data Fields

content (string): Full text of the article (title + body, concatenated with \n\n)
title (string): Article headline
category (string): Normalized topic category (54 possible values)
source (string): Original filename (e.g., asiaplus)
content_length (int64): Length of content in characters
resource (string): URL of the article (if available)
date (string): Publication date (if available; otherwise empty)

Data Splits

The dataset contains a single split (train) with all 109,082 records. Users are encouraged to create their own train/validation/test splits as needed.

Dataset Creation

Curation Rationale

The dataset was created to fill the gap in available Tajik NLP resources, particularly for unsupervised and clustering tasks. It aggregates articles from major Tajik news portals and standardizes them into a consistent format.

Source Data

Data Collection and Processing

Collection: Articles were scraped from the listed sources between 2015 and 2025. Original JSONL files contain fields such as url, title, content, category, and date.
Cleaning:
- Removed HTML tags and extra whitespace.
- Deduplicated using MD5 hashes of the content.
- Filtered out articles with content length <50 or >10,000 characters.
Category Normalization:
- Original categories (e.g., tajikistan/politics, Хабарҳо ва рӯйдодҳо, Iqtisod) were mapped to a unified set of 54 labels using a rule‑based dictionary and keyword matching.
- Empty or garbage categories (e.g., HTML‑like strings, dates) were excluded.
Final Format: Each article is stored as a JSON object with the fields described above.

Who are the source data producers?

The data were produced by the respective news portals listed in the Dataset Sources section. The content was originally created by journalists and editorial teams of those outlets.

Annotations

Categories were derived from the original category field provided by the sources, supplemented with keyword‑based rules. No manual annotation was performed; all categories are automatically normalized.

Personal and Sensitive Information

The dataset contains news articles that may mention public figures, but no attempt has been made to remove personal information (e.g., names, places). Since all data are already publicly available on the original websites, the dataset is considered non‑sensitive.

Bias, Risks, and Limitations

Source imbalance: The dataset is heavily dominated by asiaplus.tj (62% of articles), which may bias embeddings toward the style and coverage of that outlet.
Category distribution: Some categories (e.g., Ҷамъият, Иқтисод) are over‑represented, while others (e.g., Таърих, Блог) are scarce.
Date coverage: Not all articles have a reliable date field, which may affect time‑based analyses.
Language: The dataset is exclusively in Tajik (Cyrillic script). It may not generalize to other languages or scripts.

Recommendations

Users should:

Consider stratified sampling when creating splits if using the dataset for supervised tasks.
Apply appropriate weighting or subsampling to mitigate category imbalance.
Validate date information for time‑series experiments.

Citation

If you use this dataset, please cite:

@misc{arabov2025tajikcluster,
  author = {Arabov, Mullosharaf Kurbonovich},
  title = {Tajik News Cluster Dataset},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/TajikNLPWorld/tajik-news-cluster}}
}

Dataset Card Authors

Arabov Mullosharaf Kurbonovich
Email: cool.araby@gmail.com

Dataset Card Contact

For questions, please contact cool.araby@gmail.com.

Downloads last month: 12