commoncrawl_dataset / README.md
Zarinaaa's picture
create
39a0d30 verified
metadata
license: cc0-1.0
tags:
  - kyrgyz
  - common-crawl
  - web-scraping
  - text-corpus
  - low-resource-languages
  - nlp
  - turkic-languages
language:
  - ky
task_categories:
  - text-generation
  - fill-mask
  - text-classification
size_categories:
  - 100M<n<1B
pretty_name: Kyrgyz CommonCrawl Text Corpus

Kyrgyz CommonCrawl Dataset

A 271 MB text corpus of Kyrgyz language data extracted from CommonCrawl — one of the largest openly available Kyrgyz text collections for NLP research.


Dataset Description

This dataset contains Kyrgyz-language web text scraped from CommonCrawl archives, filtered by the Kyrgyz language tag (ky). The data covers a wide range of domains including news, blogs, government sites, educational content, and general web pages.

Why this matters: Kyrgyz is a low-resource Turkic language spoken by ~7 million people. High-quality text corpora are essential for training language models, yet very few large-scale Kyrgyz datasets exist publicly.


Dataset Summary

Property Value
Total size 271 MB
Language Kyrgyz (ky)
Format CSV
Source CommonCrawl (filtered by ky language tag)
Files 31 CSV files
License CC0 (public domain)

File Structure

File Size Description
data_MN.csv 29.6 MB Large text segment
data.csv 9.98 MB General web text
data_bilesinbi.csv 7.11 MB Domain-specific data
Merged file2.csv 1.66 MB Merged text segments
data_f.csv 703 kB Filtered subset
8april_final.csv 634 kB Cleaned snapshot
data.numbers 581 kB Statistics/metadata
bia.csv 37.6 kB Small subset
data_ecoproduct.csv 19.9 kB Eco/product domain
... ... Additional CSV files

Use Cases

  • Language model pretraining — Training or fine-tuning LLMs for Kyrgyz (e.g., GPT, BERT, LLaMA)
  • Text classification — Building Kyrgyz text classifiers
  • Machine translation — Source data for Kyrgyz ↔ other language pairs
  • Linguistic research — Studying modern Kyrgyz web language usage
  • Punctuation / grammar models — Training data for text normalization tools
  • NER & information extraction — Building Kyrgyz entity recognizers

Data Collection

The data was collected by:

  1. Querying CommonCrawl archives for pages tagged with the Kyrgyz language identifier (ky)
  2. Extracting text content from the matched web pages
  3. Cleaning and organizing into CSV format
  4. Deduplication and quality filtering

Preprocessing Recommendations

Before using this dataset, consider:

  • Deduplication — Web-crawled data often contains duplicate paragraphs across pages
  • Language verification — Some pages may contain mixed-language content (Kyrgyz + Russian is common)
  • Quality filtering — Remove boilerplate (navigation menus, footers, cookie notices)
  • Encoding normalization — Ensure consistent Cyrillic encoding (UTF-8)

Limitations

  • Web-crawled data may contain noise, boilerplate HTML artifacts, and mixed-language content
  • No manual curation — quality varies across files
  • Potential duplicates across different CSV files
  • Bias toward web-present content — overrepresentation of news and government text, underrepresentation of informal speech

Related Resources


Citation

@dataset{uvalieva2024kyrgyz_commoncrawl,
  author = {Uvalieva, Zarina},
  title  = {Kyrgyz CommonCrawl Text Corpus},
  year   = {2024},
  url    = {https://huggingface.co/datasets/Zarinaaa/commoncrawl_dataset}
}

Author

Zarina Uvalieva — ML Engineer specializing in NLP for low-resource languages.