salyamq/kk-corpus-v1

github repo: https://github.com/salyamq/kaz-datasets

Kazakh Corpus v1

salyamq/kk-corpus-v1 is a cleaned and deduplicated Kazakh-language corpus for language-model pretraining, tokenizer development, and Kazakh NLP research.

The corpus is derived from stukenov/sozkz-corpus-clean-v3, which is distributed under the Apache License 2.0.

The main purpose of this release is to reduce the large amount of duplicated and near-duplicated content that occurs when multiple web and text corpora are combined.

Dataset Statistics

Stage 1: cleaning statistics

Metric	Documents / characters
Original documents	13,461,211
Removed as corrupted	1,127
Removed as spam	122,240
Retained after stage 1	13,337,844
Documents changed by cleaning	4,189,524
Documents unchanged	9,147,500
Characters before cleaning	40,703,081,605
Characters after cleaning	40,609,119,108

Stage 2: deduplication statistics

The final deduplication run processed 13,314,089 documents:

Metric	Documents
Documents entering deduplication	13,314,089
Removed as near-duplicates	4,649,617
Documents retained	8,664,472
Approximate removal rate	34.9%

The final dataset contains 8,664,472 documents after cleaning, bad-character filtering, and MinHash deduplication.

Source Statistics

Source	Original	Corrupted	Spam	Retained after stage 1
madlad400	1,762,503	6	18,147	1,744,350
sib200	656	0	0	656
kazparc_sync	1,346,391	0	27	1,346,364
hplt_new	2,167,168	2	28,752	2,138,414
wikipedia	230,954	0	32	230,922
mc4	1,873,056	310	16,263	1,856,483
md_kazakhBooks	20,141	0	95	20,046
wikiann	152	0	0	152
kazsandra	39,165	0	0	39,165
md_oscar	234,943	688	3,563	230,692
culturax	2,660,249	30	40,803	2,619,416
md_kazakhNews	283,411	0	171	283,240
belebele	500	0	0	500
moscar	228,273	0	1,295	226,978
kazparc	162,817	0	11	162,806
md_leipzig	1,107,392	0	6	1,107,386
cc100	1,343,150	91	13,077	1,329,982

Cleaning Pipeline

The initial cleaning stage performs the following operations:

1. Decodes escaped line breaks, tabs, carriage returns, and common HTML entities.
2. Removes paragraphs containing the Unicode replacement character �.
3. Removes empty paragraphs.
4. Normalizes mixed Cyrillic and Latin homoglyphs.
5. Collapses excessive repeated punctuation.
6. Removes HTML tags and invisible Unicode characters.
7. Applies Unicode NFC normalization and whitespace cleanup.
8. Removes documents detected by spam heuristics.

Deduplication

Deduplication is the central step of this release. The source corpus combines many datasets that often contain the same articles, web pages, boilerplate, or lightly modified copies. Keeping all such copies would artificially inflate the corpus and cause a language model to over-weight repeated text.

The pipeline uses the DataTrove MinHash deduplication workflow:

1. Remove documents containing a predefined set of problematic Unicode characters.
2. Build MinHash signatures using 5-gram shingles.
3. Compare documents using 16 buckets and 8 hashes per bucket.
4. Cluster documents with similar MinHash signatures.
5. Keep one representative document from each duplicate or near-duplicate cluster.

This removes both exact duplicates and documents that are highly similar, including copied articles with small changes in punctuation, whitespace, or formatting. The deduplication stage removed 4,649,617 documents, or approximately 34.9% of the documents entering that stage.

Schema

A typical record contains:

• text — cleaned document text;
• source — original dataset source, when available.

Example:

{
  "text": "Қазақстан туралы мәтін...",
  "source": "culturax"
}

Intended Uses

This corpus can be used for:

• pretraining Kazakh-language language models;
• continued pretraining;
• tokenizer development;
• language identification;
• text normalization research;
• corpus-quality analysis;
• Kazakh NLP experiments.

Citation

If you use this dataset, please cite it as:

@misc{salyamq_kk_corpus_v1,
  author       = {{salyamq}},
  title        = {Kazakh Corpus v1},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/salyamq/kk-corpus-v1}}
}

License

This dataset is released under the Apache License 2.0. See the Apache License 2.0 for details.