A comprehensive multilabel classification dataset for categorizing Tatar language news articles into multiple topics, curated by TatarNLPWorld as part of the Tat2Vec project.

📖 Overview

This dataset contains 55,709 Tatar language texts classified into 13 distinct labels. Each entry includes the full content, title, labels (multi-hot encoding), label vector, original category, source, and publication date. It's specifically designed for training and evaluating multilabel text classification models for Tatar language natural language processing tasks.

📊 Dataset Statistics

General Information

Metric	Value
Total Records	55,709
Records with Title	55,709 (100%)
Records with Date	55,709 (100%)
Total Characters	132,158,076
Average Content Length	2,372 characters
Median Content Length	1,189 characters
Content Length Range	50 - 10,000 characters
Date Range	2008 - 2026
Number of Labels	13
Average Labels per Sample	3.72

Label Distribution

Label (Tatar)	Label (English)	Count	Percentage
Татарстан	Tatarstan	35,943	64.5%
Җәмгыять	Society	32,896	59.0%
Дин	Religion	21,845	39.2%
Дөнья	World	18,281	32.8%
Мәгариф	Education	16,231	29.1%
Икътисад	Economy	16,032	28.8%
Мәдәният	Culture	14,585	26.2%
Милләт	Nation	12,163	21.8%
Шоу-бизнес	Show Business	10,120	18.2%
Сәламәтлек	Health	8,585	15.4%
Сәясәт	Politics	8,520	15.3%
Әдәбият	Literature	6,458	11.6%
Спорт	Sports	5,666	10.2%

Original Categories Distribution (Top 10)

Original Category	Count	Percentage
Җәмгыять	8,053	16.1%
Мәдәният	5,987	11.9%
Шоу-бизнес	5,773	11.5%
Яңалыклар	4,229	8.4%
Дин	1,930	3.8%
Архив	1,316	2.6%
Тормыш сулышы	1,019	2.0%
Гыйбрәт ал	959	1.9%
Мәгариф	956	1.9%
Милләт	941	1.9%

📁 Data Structure

Each record contains the following fields:

Field	Type	Description
`content`	string	Full article content in Tatar language
`title`	string	Article title
`labels`	list[int]	Label indices for multi-hot encoding
`label_vector`	list[int]	Multi-hot vector representation (13 dimensions)
`num_labels`	integer	Number of labels assigned to this article
`original_category`	string	Original single-category classification
`content_length`	integer	Length of content in characters
`resource`	string	Source URL or identifier
`date`	string	Publication date

🚀 Usage Example

Loading the Dataset

from datasets import load_dataset
from collections import Counter
import numpy as np

# Load the dataset
dataset = load_dataset("TatarNLPWorld/tatar-news-analysis-multilabel")

# Access train and validation splits
train_data = dataset["train"]
validation_data = dataset["validation"]

print(f"Training samples: {len(train_data)}")
print(f"Validation samples: {len(validation_data)}")
print(f"Number of labels: {len(set([l for labels in train_data['labels'] for l in labels]))}")
print(f"Labels per sample (avg): {np.mean([len(l) for l in train_data['labels']]):.2f}")
print(f"Content length (avg): {np.mean(train_data['content_length']):.0f} characters")

# Check label distribution
all_labels = [label for labels in train_data['labels_text'] for label in labels]
distribution = Counter(all_labels)
print("\n### Label Distribution")
for label, count in distribution.most_common():
    print(f"- **{label}**: {count} ({count/len(train_data)*100:.1f}%)")

# Check original categories
orig_cats = Counter(train_data['original_category'])
print("\n### Original Categories (Top 10)")
for cat, count in orig_cats.most_common(10):
    print(f"- **{cat}**: {count} ({count/len(train_data)*100:.1f}%)")

Expected Output:

Training samples: 50138
Validation samples: 5571
Number of labels: 13
Labels per sample (avg): 3.72
Content length (avg): 2370 characters

### Label Distribution
- **Татарстан**: 32289 (64.4%)
- **Җәмгыять**: 29553 (58.9%)
- **Дин**: 19622 (39.1%)
- **Дөнья**: 16463 (32.8%)
- **Мәгариф**: 14578 (29.1%)
- **Икътисад**: 14448 (28.8%)
- **Мәдәният**: 13113 (26.2%)
- **Милләт**: 10903 (21.7%)
- **Шоу-бизнес**: 9088 (18.1%)
- **Сәламәтлек**: 7710 (15.4%)
- **Сәясәт**: 7621 (15.2%)
- **Әдәбият**: 5788 (11.5%)
- **Спорт**: 5092 (10.2%)

### Original Categories (Top 10)
- **Җәмгыять**: 8053 (16.1%)
- **Мәдәният**: 5987 (11.9%)
- **Шоу-бизнес**: 5773 (11.5%)
- **Яңалыклар**: 4229 (8.4%)
- **Дин**: 1930 (3.8%)
- **Архив**: 1316 (2.6%)
- **Тормыш сулышы**: 1019 (2.0%)
- **Гыйбрәт ал**: 959 (1.9%)
- **Мәгариф**: 956 (1.9%)
- **Милләт**: 941 (1.9%)

Sample Record

{
    "content": "Шактый катлаулы көннәрдә яшибез. Сугышлар, кризис, доллар курсы – болар барысы да көн саен һәр тарафтан ишетелә, иң оптимист кешене дә шөбһәгә сала торган сүзләргә әйләнде. Иртән ике баланы кайдан булуым белән горурлану хисләре туды. Бик зур рәхмәт сезгә, Актаныштагы яңа танышларым!",
    "title": "«Агыйдел»гә таң калдым",
    "labels": ["Дөнья", "Шоу-бизнес", "Татарстан", "Җәмгыять"],
    "label_vector": [1, 1, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0],
    "num_labels": 4,
    "original_category": "Аулак өй",
    "content_length": 4526,
    "resource": "https://beznen.ru/archive/aulak-oy/261215/agyydelg-tan-kaldym",
    "date": "2015-12-26T00:00:00"
}

🔬 Applications

This dataset supports multiple NLP tasks:

Multi-label Text Classification: Topic categorization of Tatar news with multiple labels per article
Token Classification: Named entity recognition, part-of-speech tagging
Feature Extraction: Creating embeddings for Tatar language texts
Sentence Similarity: Comparing semantic similarity between texts
Text Generation: Language modeling and text generation tasks
Summarization: Generating summaries of Tatar news/articles
Zero-shot Classification: Cross-lingual and zero-shot learning experiments

📊 Dataset Splits

The dataset is automatically stratified split into:

Training Set: 90% (50,138 samples)
Validation Set: 10% (5,571 samples)

The splits maintain the original label distribution.

🛠️ Data Collection

This dataset is part of the Tat2Vec project and was curated from various Tatar language news sources and publications spanning from 2008 to 2026, ensuring:

Authentic Tatar language content
Diverse topics across 13 labels
Large-scale coverage with nearly 56,000 samples
Wide temporal diversity
Multi-label annotations averaging 3.72 labels per article

📜 License

This dataset is released under the MIT License, allowing for:

Commercial use
Modification
Distribution
Private use

🤝 Citation

If you use this dataset in your research or projects, please cite:

@dataset{tat2vec_multilabel_2026,
    title = {Tatar News Multilabel Classification Dataset},
    author = {TatarNLPWorld},
    year = {2026},
    publisher = {Hugging Face},
    version = {1.0.0},
    note = {Part of the Tat2Vec project},
    url = {https://huggingface.co/datasets/TatarNLPWorld/tatar-news-analysis-multilabel}
}

👥 Team and Maintenance

This dataset is maintained by TatarNLPWorld, a community dedicated to advancing Natural Language Processing for the Tatar language through open-source resources and collaboration. It is part of the larger Tat2Vec project initiative.

Contributors

TatarNLPWorld Team

📬 Contact and Contributions

We welcome contributions and feedback!

Issues: Please open an issue on the Hugging Face repository
Contributions: Submit pull requests for improvements
Contact: Reach out through the TatarNLPWorld community channels

🌟 Acknowledgments

Special thanks to all data sources and contributors who made this dataset possible, supporting the development of Tatar language NLP resources through the Tat2Vec project.

Downloads last month: 15