TatarNLPWorld/vk-groups

VK Groups Dataset: Tatar Social Media Posts

Dataset Description

This dataset contains posts from five popular VKontakte (VK) communities targeting Tatar-speaking audiences. It includes content across different themes: religion, relationships, humor, advice, and cooking. The dataset is designed for various NLP tasks including text classification, sentiment analysis, topic modeling, and engagement prediction.

• Curated by: TatarNLPWorld
• Language(s): Tatar (Cyrillic script), some Russian code-switching
• License: MIT

Dataset Sources

• Repository: https://huggingface.co/datasets/TatarNLPWorld/vk-groups
• Source VK Groups:
  • Религия (Religion): https://vk.com/allahtan
  • Ялгыз Тушәк (Lonely Bed/Relationships): https://vk.com/ialgyz16
  • Сазлык (Swamp/Humor): https://vk.com/public147862424
  • КИҢӘШ КИРӘК (Advice Needed): https://vk.com/sovet_kinesh
  • Кулинария (Cooking): https://vk.com/public221151764

Dataset Summary

Statistic	Value
Total posts	56,805
Total classes	5
Date range	2015-06-17 to 2026-03-11
Total likes	2,708,506
Total views	344,348,797
Average engagement rate	40.17%

Class Distribution

Class	Count	Percentage
Сазлык (Humor)	22,040	38.8%
Религия (Religion)	13,017	22.9%
КИҢӘШ КИРӘК (Advice)	10,187	17.9%
Ялгыз Тушәк (Relationships)	9,688	17.1%
Кулинария (Cooking)	1,873	3.3%

Dataset Structure

Two configurations are available:

1. full - Complete dataset with all features

• text: Post content (string)
• class: Class name in Tatar (string)
• source_url: URL of the source VK group (string)
• source_type: Always "vk_group" (string)
• likes: Number of likes (int64)
• comments: Number of comments (int64)
• reposts: Number of reposts (int64)
• views: Number of views (int64)
• date: Post publication date (string)
• post_id: Unique post identifier (int64)
• label: Numeric class label (0-4) (int64)
• engagement_total: Total interactions (likes+comments+reposts) (int64)
• engagement_rate: (engagement_total / views) * 100 (float64)

2. simple - Simplified version for classification

• text: Post content (string)
• class: Class name in Tatar (string)
• label: Numeric class label (0-4) (int64)

Data Splits

All configurations are split into:

Split	Size	Percentage
Train	45,444	80%
Validation	5,680	10%
Test	5,681	10%
Full	56,805	100%
Sample	1,000	1.8%

Dataset Creation

Data Collection

Posts were collected from five public VK communities using the VK API. The data includes all publicly available posts from each group's inception through March 2026.

Data Processing

1. Raw posts were extracted with metadata (likes, comments, reposts, views, date)
2. Empty or invalid posts (length < 10 characters) were removed
3. Duplicate posts were deduplicated
4. Posts consisting only of special characters/links were filtered out
5. Final dataset: 56,805 posts (from original 117,831, 48.2% retained)

Label Mapping

Label	Class
0	КИҢӘШ КИРӘК (Advice)
1	Кулинария (Cooking)
2	Религия (Religion)
3	Сазлык (Humor)
4	Ялгыз Тушәк (Relationships)

Uses

Direct Use

• Text Classification: Multi-class classification of Tatar social media content
• Sentiment Analysis: Analyze emotional tone across different community types
• Topic Modeling: Discover themes in Tatar-language social media
• Engagement Prediction: Predict post popularity based on content
• Code-switching Analysis: Study Tatar-Russian language mixing
• Cultural Studies: Research Tatar online communities and discourse

Out-of-Scope Use

• Not suitable for training production systems without careful bias evaluation
• Not representative of all Tatar speakers or online communities
• Contains user-generated content with potential biases

Bias, Risks, and Limitations

1. Class imbalance: Cooking class (3.3%) is significantly underrepresented
2. Platform bias: Only VKontakte, may not represent other social media
3. Language variation: Mix of Tatar and Russian may affect model performance
4. Temporal bias: 10+ years of data, language use may have evolved
5. Content bias: Topics reflect community interests, not general Tatar discourse

Citation

If you use this dataset, please cite:

@dataset{vk_groups_tatar_2026, title = {VK Groups Dataset: Tatar Social Media Posts}, author = {Arabov, Mullosharaf Kurbonovich and TatarNLPWorld}, year = {2026}, publisher = {Hugging Face}, url = {https://huggingface.co/datasets/TatarNLPWorld/vk-groups}, note = {Created by Arabov Mullosharaf Kurbonovich for TatarNLPWorld organization} }

How to Use

from datasets import load_dataset

# Load full dataset
dataset_full = load_dataset('TatarNLPWorld/vk-groups', 'full')
train_full = dataset_full['train']

# Load simple dataset (for classification)
dataset_simple = load_dataset('TatarNLPWorld/vk-groups', 'simple')
train_simple = dataset_simple['train']

# Access data
texts = train_simple['text']
labels = train_simple['label']
classes = train_simple['class']

Contact

For questions or collaborations, please contact TatarNLPWorld.

---

Languages: Tatar (Cyrillic), some Russian
License: MIT
Size: 56,805 posts
Last Updated: March 2026