📚 Tatar Wiki Corpus (tatar-wiki-corpus)

A comprehensive cleaned corpus of Tatar Wikipedia and Wikibooks with over 467,000 articles. This dataset is ideal for training language models, text classification, information retrieval, and various NLP tasks for the Tatar language.

---

📊 Corpus Statistics

Overall Statistics

Metric	Value
📄 Documents	467,578
📝 Characters	336,591,416
🔤 Words	47,742,463
📏 Sentences	8,547,916
📚 Unique words	830,915
🎯 Unique characters	152
📏 Avg. document length	720 characters
📏 Avg. word length	6.06 characters
📏 Avg. sentence length	5.59 words
🧹 Compression after cleaning	27.3%

📁 Source Distribution

Source	Documents	Percentage
Wikipedia (wikipedia)	467,013	99.88%
Wikibooks (wikibooks)	565	0.12%

🔤 Most Frequent Words (Cleaned)

Word	Frequency	Translation
торак	879,185	settlement / dwelling
урнашкан	652,623	located
һәм	538,061	and
буенча	500,623	according to / along
төркем	430,100	category / group
пунктларытөркем	357,180	category of settlements
әлифба	333,379	alphabet (by letter)
уртача	269,028	average
ред	232,450	editor
пунктлар	197,878	settlements
авыл	176,660	village
диңгез	150,056	sea
елга	143,379	river
башкаласы	141,272	capital city
халык	113,915	people / population
су	105,129	water
энциклопедия	121,355	encyclopedia

🔤 Most Frequent Characters

Character	Frequency
space	47,119,926
а	27,887,729
р	18,317,042
е	17,511,687
н	17,048,091
к	14,982,961
л	13,713,831
и	12,868,963
т	12,689,023
ы	11,161,694

🏷️ Top Categories

Category	Documents
Әлифба буенча торак пунктлар (Settlements by alphabet)	324,459
Мексика торак пунктлары (Mexican settlements)	55,083
Италия торак пунктлары (Italian settlements)	45,150
Польша торак пунктлары (Polish settlements)	42,949
Төркия торак пунктлары (Turkish settlements)	39,614
Франция коммуналары (French communes)	35,805
Беларусия авыллары (Belarusian villages)	23,287
Төркия мәхәлләләре (Turkish neighborhoods)	21,597
Веракрус торак пунктлары (Veracruz settlements)	21,419
АКШ шәһәрләре (US cities)	19,351

---

📌 Record Structure

Each line in the JSONL file contains the following fields:

{
  "title": "Әгерҗе",
  "text": "Әгерҗе - РФнең субъекты булган Татарстан Республикасындагы шәһәр...",
  "categories": ["Татарстан район үзәкләре", "Әгерҗе районының торак пунктлары"],
  "alphabet": "cyrillic",
  "alphabet_info": {
    "alphabet": "cyrillic",
    "confidence": 0.9369,
    "cyr_ratio": 0.9369,
    "lat_ratio": 0.0631
  },
  "quality_info": {
    "word_count": 263,
    "sentence_count": 20,
    "readability_score": 93.7,
    "avg_sentence_length": 13.15
  },
  "text_length": 1969,
  "categories_count": 2,
  "url": "https://tt.wikipedia.org/wiki/Әгерҗе",
  "source": "wikipedia"
}

Field Descriptions

Field	Type	Description
title	string	Article title
text	string	Cleaned text of the article (no HTML, no empty brackets)
categories	list of strings	Wikipedia categories associated with the article
alphabet	string	Detected alphabet (cyrillic)
alphabet_info	dict	Detailed information about alphabet detection
quality_info	dict	Text quality metrics (word count, readability, etc.)
text_length	int	Length of text in characters
categories_count	int	Number of categories
url	string	Original Wikipedia/Wikibooks URL
source	string	Source type: wikipedia or wikibooks

Data Quality Features

✅ Fully cleaned text - All HTML tags, URLs, emails, and special characters removed
✅ Empty brackets removed - Patterns like (), (..), (1), (*) are cleaned
✅ Language detection - Cyrillic ratio and confidence scores included
✅ Quality metrics - Word count, sentence count, readability score
✅ Original metadata - Categories, URLs, and source attribution preserved

---

🚀 Loading the Dataset

Using Hugging Face datasets (recommended)

from datasets import load_dataset

# Load full corpus
dataset = load_dataset("TatarNLPWorld/tatar-wiki-corpus", split="full")
print(f"Total documents: {len(dataset)}")  # 467,578

# View first record
print(dataset[0])

# Access specific fields
titles = dataset["title"]
texts = dataset["text"]
categories = dataset["categories"]
sources = dataset["source"]

Available Splits

• full — complete corpus (467,578 documents)
• sample — small subset for testing (1,000 documents)

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

# Load specific split
full = load_dataset("TatarNLPWorld/tatar-wiki-corpus", split="full")

Quick Start Example

# 1️⃣ Install if needed
!pip install -q datasets

# 2️⃣ Load dataset
from datasets import load_dataset

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

# 3️⃣ Check basic info
print(f"Total documents: {len(dataset)}")
print(f"Available fields: {dataset.column_names}")

# 4️⃣ View sample
print("Sample record:")
print(dataset[0])

# 5️⃣ Access cleaned text
first_text = dataset[0]["text"]
print(f"First 500 chars:\n{first_text[:500]}")

# 6️⃣ Filter by source
wikipedia = dataset.filter(lambda x: x["source"] == "wikipedia")
wikibooks = dataset.filter(lambda x: x["source"] == "wikibooks")

print(f"Wikipedia articles: {len(wikipedia)}")
print(f"Wikibooks articles: {len(wikibooks)}")

---

🎯 Recommended Usage

Task	What to Use
Language Models (GPT, BERT, LLaMA)	text field (cleaned article text)
Text Classification	text + categories
Information Retrieval	text field + embeddings
Title Generation	text → title
Source Attribution	source field (wikipedia/wikibooks)
Word Embeddings (Word2vec, fastText)	text field (tokenized)
Readability Analysis	quality_info metrics
Question Answering	text field as context
Summarization	text → title or extractive summaries

Examples for Different Tasks

# For language model training
train_texts = dataset["text"]

# For multi-label classification with categories
texts = dataset["text"]
labels = dataset["categories"]  # Each document can have multiple categories

# For article retrieval
from sentence_transformers import SentenceTransformer
model = SentenceTransformer('intfloat/multilingual-e5-small')
embeddings = model.encode(dataset["text"][:1000])  # Encode first 1000 articles

# Filter by source
wikipedia_only = dataset.filter(lambda x: x["source"] == "wikipedia")

# Filter by text quality
high_quality = dataset.filter(lambda x: x["quality_info"]["word_count"] > 100)

# Get documents with specific categories
geography_articles = dataset.filter(
    lambda x: any("торак пунктлар" in cat for cat in x["categories"])
)

---

📝 Sample Records

Record 1 (Wikipedia - City):

• Title: "Әгерҗе"
• Source: wikipedia
• Categories: ["Татарстан район үзәкләре", "Әгерҗе районының торак пунктлары"]
• Text excerpt: "Әгерҗе - РФнең субъекты булган Татарстан Республикасындагы шәһәр, Әгерҗе районының үзәге. Мөһим тимер юл төене. Халык саны – 19 739 кеше (2016)..."

Record 2 (Wikipedia - River):

• Title: "Идел"
• Source: wikipedia
• Categories: ["Россия елгалары", "Идел"]
• Text excerpt: "Идел - Европаның иң озын елгасы. Озынлыгы 3530 км, бассейнының мәйданы 1360 мең км²..."

Record 3 (Wikibooks - Folk Tale):

• Title: "Тапкыр солдат"
• Source: wikibooks
• Categories: ["Татар халык әкиятләре"]
• Text excerpt: "Тапкыр солдат (татар халык әкияте) Унике ел патша армиясендә хезмәт иткән солдатның өенә кайтыр вакыты җиткән..."

---

⚖️ License

This dataset is distributed under the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license, consistent with Wikipedia's licensing.

Terms of Use:

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

Source Attribution:

All texts originate from Tatar Wikipedia (tt.wikipedia.org) and Tatar Wikibooks (tt.wikibooks.org). Original URLs are preserved in the url field.

---

🛡️ Privacy & Safety

The dataset has been automatically cleaned of:

• HTML tags and markup
• URLs and email addresses
• Phone numbers
• Copyright notices
• Empty brackets and formatting artifacts
• Non-Tatar characters (optional, configurable)

All texts are sourced from publicly available Wikimedia projects and have been processed for NLP use. No personal or sensitive information is intentionally included.

---

📊 Cleaning Process

The corpus underwent extensive cleaning using a custom Tatar text cleaner:

1. HTML decoding and tag removal
2. URL and email removal
3. Copyright notice removal
4. Empty bracket removal - (), (..), (...), (1), (*)
5. Repeated punctuation normalization
6. Special character cleaning
7. Non-Tatar character removal (configurable)
8. Space and line normalization
9. Quality filtering (minimum length, Tatar ratio, etc.)

Result: 27.3% size reduction while preserving all meaningful content.

---

📖 Citation

@misc{tatarwikicorpus2026,
  author = {TatarNLPWorld},
  title = {Tatar Wiki Corpus: Cleaned Tatar Wikipedia and Wikibooks Dataset},
  year = {2026},
  month = {March},
  note = {Version with 467,578 cleaned documents},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/TatarNLPWorld/tatar-wiki-corpus}}
}

BibTeX for academic papers

@dataset{tatar_wiki_corpus_2026,
  title={Tatar Wiki Corpus: A Large-Scale Cleaned Dataset of Tatar Wikipedia and Wikibooks},
  author={{TatarNLPWorld}},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/TatarNLPWorld/tatar-wiki-corpus}
}

---

📬 Contact & Community

• Organization: TatarNLPWorld on Hugging Face
• Questions & Feedback: Open an issue or discussion
• Contributions: Welcome! If you have additional Tatar texts to contribute, please contact us

Join the Tatar NLP Community

• Share your models trained on this corpus
• Report issues or suggest improvements
• Contribute to Tatar language technology development

---

🙏 Acknowledgments

Special thanks to:

• Wikimedia Foundation for Tatar Wikipedia and Wikibooks content
• All volunteer contributors to Tatar Wikimedia projects
• The global NLP community for tools and inspiration
• Everyone working to preserve and promote the Tatar language

Data Sources

• Wikipedia Tatar (tt.wikipedia.org) - Main source (467,013 articles)
• Wikibooks Tatar (tt.wikibooks.org) - Supplementary source (565 books/folios)

---

📋 Changelog

Version 1.0 (2026-03-03)

• Initial release
• 467,578 cleaned documents
• 877 MB total size
• Wikipedia + Wikibooks sources
• Full metadata preservation
• 27.3% size reduction after cleaning

---

Made with ❤️ for the Tatar language and the global NLP community
Без татар теле өчен эшлибез 🇹🇹