culturay-kk-v1

A dataset of Kazakh-language texts derived from CulturaY — a large multilingual corpus built from Internet Archive data. The kk subset went through an additional multi-stage cleaning and quality-filtering pipeline. Please note that this dataset is the first version and may contain mixed Russian-Kazakh text, minor spam, and adult content. This dataset is already ready for SLM / LLM training; however, we are currently preparing a large Kazakh corpus that will be several hundred times larger than this one and cleaner, where training will be highly effective.

Overview

This dataset contains cleaned Kazakh-language documents that have passed through a pipeline of filtering, normalization, PII removal, and deduplication.

Parameter	Value
Language	Kazakh (kk)
Source	CulturaY → Internet Archive
Final size	197,937 documents
License	Apache 2.0

Description	Count	Percentage
Initial Documents	274,770	100.00%
Current Documents	197,937	72.04%
Documents Removed	76,833	27.96%

Data Structure

Column	Description
text	Document text
collection	Source collection (original origin of the document)
resource	Dataset source (always CulturaY)

Cleaning Pipeline

Below are the cleaning stages in the order they were applied, along with the available statistics for each step.

#	Stage	Description	Statistics
1	ftfy initialization	Fix encoding artifacts (e.g., ’ → ') before any structural string manipulation	Records fixed: 53,060
2	HTML extraction	Strip HTML tags, JavaScript, and navigation elements using trafilatura, significantly reducing payload size	—
3	Unicode normalization (NFC)	Normalize text for consistent representation of Cyrillic characters	Broken records: 0
4	Control character cleanup	Remove corrupted or extraneous control characters after normalization	Broken records: 0
5	Language validation & filtering	Lightweight check for at least one Kazakh Cyrillic character + language identification via fastText (confidence threshold ≥ 0.65), cutting CPU usage by up to 80% through early filtering	Lines dropped: 2,764,786
6	PII removal	Strip emails, IP addresses, phone numbers, and IINs (12-digit national IDs)	Replacements made: 45,650
7	URL density filtering	Discard documents with URL density > 5 per 1,000 characters (spam / low-quality content)	260,096 → 260,053 (deleted 43)
8	Whitespace collapsing	Normalize whitespace by collapsing redundant spaces and line breaks left over from HTML stripping and PII redaction	—
9	Length filtering	Discard documents shorter than 25 characters (word counts computed on the final cleaned text)	14,357,005 → 12,734,908 (deleted 1,622,097)
10	Heuristic filtering (Gopher rules)	At least 70% letter characters, < 30% of lines ending in ellipsis/punctuation, average sentence length of 3–100 words	260,053 → 248,625 (deleted 11,428)
11	C4 quality filtering	Discard boilerplate / low-quality content (e.g., "Lorem Ipsum", "JavaScript required", cookie policy notices, and their Kazakh equivalents)	248,625 → 248,548 (deleted 77)
12	Gzip repetition filter	Discard documents with excessive repetitive, low-information text (compression-ratio threshold ≥ 0.20)	248,548 → 247,779 (deleted 769)
13	Exact deduplication	Remove exact duplicates	implemented
14	Near-duplicate detection (MinHash LSH)	128 permutations, Jaccard similarity threshold ≥ 0.9, removing rephrased duplicates	247,779 → 197,937 (deleted 49,842)

Note on units: steps 5, 6, and 9 operate at the line/segment level (hence the much larger numbers, in the millions), while steps 7 and 10–14 operate at the document level. The final dataset size after all stages is 207,039 documents.

License

This dataset is distributed under the Apache 2.0 license.

Additional

• Repo: github.com/salyamq/dataset-culturay-v1
• Contact: salim09291610@gmail.com / s_tokhtobayev@kbtu.kz

Usage

from datasets import load_dataset

ds = load_dataset("salyamq/culturay-kk-v1")
print(ds["train"][0])

Citation

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

@dataset{tokhtobayev2026culturaykk,
  author = {Salim Tokhtobayev},
  title = {Culturay-KK-v1: A Kazakh Language Text Corpus},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/salyamq/culturay-kk-v1}
}