📚 Tatar Web Corpus (tatar-web-corpus)

The largest open corpus for the Tatar language with over 1 million documents collected from news websites, social media, articles, books, and Wikipedia. Designed for various NLP tasks including language modeling, text classification, information extraction, and search.

---

📊 Corpus Statistics

Overall Statistics

Metric	Value
📄 Documents	1,053,698
📝 Characters	1,796,654,320
🔤 Words	254,462,180
🎯 Tokens	304,923,850
📚 Unique words (approx.)	1,258,563
📏 Avg. document length	1,705 characters
📌 With title	1,035,160 (98.2%)
📌 With category	1,032,024 (97.9%)
📌 With source	1,053,240 (100%)

📁 Source Distribution

Source	Documents	Percentage
ttwiki_cyrillic	853,196	80.97%
merged_matbugat_news	56,558	5.37%
intertat_news	55,008	5.22%
beznen_articles	24,884	2.36%
merged_azatliqorg	16,632	1.58%
vk_posts_filtered	12,248	1.16%
syuyumbike_news	8,192	0.78%
shahrikazan_news	6,446	0.61%
tatar_inform	5,144	0.49%
kazanutlary_articles	4,038	0.38%
alluki_articles	3,928	0.37%
merged_mamadysh_tt	3,444	0.33%
webbooks	1,384	0.13%
tuganaylar_articles	680	0.06%
vatantat_news	490	0.05%
telegram	458	0.04%
nurlat_articles_tt_only_soft	338	0.03%
kiziltan_articles	280	0.03%
books	194	0.02%
belgechbook	92	0.01%
almet_articles	64	0.01%

Total sources: 21

🔤 Most Frequent Words

Word	Frequency
да	209,407
белән	158,340
ул	156,740
дә	149,821
һәм	137,386
дип	120,781
бу	116,989
иде	107,990
бер	95,129
мин	76,308
өчен	72,522
инде	60,870
бик	58,640
аның	57,926
бар	51,666
генә	49,173
гына	47,670
шул	47,393
юк	43,893
түгел	43,491

---

📌 Record Structure

Each line in the JSONL file contains the following fields:

{
  "title": "Document title",
  "content": "Main text...",
  "category": "News / Blog / Book / Wikipedia",
  "source": "Source URL or channel name",
  "text": "Document title\n\nMain text"  // Convenience field for modeling
}

Field Descriptions

Field	Description	Availability
title	Document title	98.2%
content	Main text	100%
category	Topic category	97.9%
source	Source URL or channel name	100%
text	Ready-to-use field for modeling (title + \n\n + content)	100%

🔗 Source Attribution

• Wikipedia articles: tt.wikipedia.org
• Books: Telegram channels (t.me/tatarkit, t.me/tatarkitaphane, t.me/tatelkit, t.me/tatarkitap)
• News articles: Original URLs when available

---

🚀 Loading the Dataset

Using Hugging Face datasets (recommended)

from datasets import load_dataset

# Load full corpus
dataset = load_dataset("TatarNLPWorld/tatar-web-corpus", split="full")
print(f"Total documents: {len(dataset)}")  # 1,053,698

# View first record
print(dataset[0])

# Use 'text' field for modeling
texts = dataset["text"]

Available Splits

• full — complete corpus (1,053,698 documents)
• sample — small subset for testing (1,000 documents)

# Load sample split
sample = load_dataset("TatarNLPWorld/tatar-web-corpus", split="sample")

Quick Start Example

# 1️⃣ Install library if needed
!pip install -q datasets

# 2️⃣ Import and load
from datasets import load_dataset

dataset = load_dataset("TatarNLPWorld/tatar-web-corpus", split="full")

# 3️⃣ Check basic info
print(f"Total documents: {len(dataset)}")
print("Sample record:")
print(dataset[0])

# 4️⃣ Access 'text' field
texts = dataset["text"]
print(f"First 500 chars of first document:\n{texts[0][:500]}")

---

🎯 Recommended Usage

Task	What to Use
Language Models (GPT, BERT)	text field (title + content)
Text Classification	text + category
Title Generation	content → title
Source Analysis	source field
Word Embeddings (Word2vec, fastText)	text field

Examples for Different Tasks

# For GPT / language modeling
train_texts = dataset["text"]

# For classification
train_texts = dataset["text"]
train_labels = dataset["category"]

# For summarization
articles = dataset["content"]
titles = dataset["title"]

# Filter by category
news_only = dataset.filter(lambda x: x["category"] == "яңалыклар")
wikipedia_only = dataset.filter(lambda x: x["source"] == "ttwiki_cyrillic")

---

📝 Sample Records

Record 1 (News):

• Title: "Татарстан Республикасы Президенты Рөстәм Миңнехановның мөрәҗәгате"
• Category: яңалыклар (news)
• Source: Intertat
• Content: «Хөрмәтле татарстанлылар! Кадерле дуслар!..»

Record 2 (Wikipedia):

• Title: "Казан"
• Category: Wikipedia
• Source: ttwiki_cyrillic
• Content: «Казан - Татарстанның башкаласы, Идел буендагы борынгы шәһәрләрнең берсе...»

Record 3 (Fiction):

• Title: "Кышкы урманда"
• Category: нәфис әдәбият (fiction)
• Source: t.me/tatarkit
• Content: «Урман тынлыгында кар ява иде...»

---

⚖️ License

This dataset is distributed under the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license.

Terms of Use:

• Attribution — You must give appropriate credit to TatarNLPWorld
• ShareAlike — If you modify the material, distribute under the same license
• Commercial use is allowed

Source Attribution:

Texts are collected from publicly available sources. Each document links to its original source in the source field.

---

🛡️ Privacy & Safety

The dataset has been cleaned of personal information (PII). However, if you find any sensitive data, please report it.

---

📖 Citation

@misc{tatarwebcorpus2026,
  author = {TatarNLPWorld},
  title = {Tatar Web Corpus: A Large-Scale Dataset for Tatar Language NLP},
  year = {2026},
  month = {February},
  note = {Version with 1,053,698 documents},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/TatarNLPWorld/tatar-web-corpus}}
}

---

📬 Contact

• Organization: TatarNLPWorld on Hugging Face
• Questions: Open an issue

---

🙏 Acknowledgments

Special thanks to all data sources:

• Wikipedia (tt.wikipedia.org)
• News portals: Intertat, Tatar-inform, Shähri Qazan, Vatanym Tatarstan
• Magazines: Söyembikä, Qazan utları
• Social media: VK thematic groups
• Books: Tatar books Telegram channels

---

Made with ❤️ for the Tatar language and the global NLP community