stukenov/sozkz-corpus-clean-v3

SozKZ Corpus Clean v3 — Cleaned Kazakh Text Corpus

A large-scale cleaned and deduplicated Kazakh text corpus assembled from 18 public sources. Designed for pre-training causal language models on Kazakh text.

Overview

Total texts	13,700,018
Train split	~13,563,018 (99%)
Validation split	~137,000 (1%)
Raw input	28,431,116 texts
Pass rate	48.2%
Dedup removed	3,170,330
License	Apache 2.0

Sources

Source	Raw	Clean	Pass %	Type
culturax	2,731,934	2,707,214	99.1%	Web crawl (CulturaX)
madlad400	1,807,996	1,794,308	99.2%	Web crawl (MADLAD-400)
hplt_new	2,637,330	2,204,165	83.6%	Web crawl (HPLT)
mc4	2,371,528	1,906,763	80.4%	Web crawl (mC4)
cc100	1,721,481	1,365,739	79.3%	Web crawl (CC-100)
kazparc_sync	9,632,030	1,370,243	14.2%	Parallel corpus (KazParC)
md_leipzig	1,706,485	1,128,122	66.1%	Leipzig corpora collection
md_kazakhNews	3,264,273	288,247	8.8%	Kazakh news articles
md_oscar	269,047	239,807	89.1%	OSCAR web corpus
moscar	245,869	231,693	94.2%	OSCAR (secondary)
kazparc	1,647,560	165,754	10.1%	Parallel corpus (KazParC)
kazsandra	146,253	40,236	27.5%	Kazsandra corpus
md_kazakhBooks	8,423	20,482	—	Kazakh literary texts*
wikipedia	~20,000	~19,988	~99%	Kazakh Wikipedia
belebele	900	488	54.2%	Belebele benchmark
sib200	701	648	92.4%	SIB-200 benchmark
wikiann	~1,000	~966	—	WikiANN NER

*kazakhBooks raw count < clean count because long texts are split into chunks during processing.

Data Sources

• kz-transformers/multidomain-kazakh-dataset: oscar, kazakhNews, kazakhBooks, leipzig CSVs
• Collected parquets (wave 1): culturax, hplt_new, madlad400, mc4, cc100, kazparc, kazparc_sync, moscar, wikipedia, belebele, sib200, wikiann, kazsandra
• Collected parquets (wave 2): additional sources

Cleaning Pipeline

Nine-stage filter pipeline, ordered from fast to slow:

#	Filter	Description	Threshold
1	OSCAR dict fix	Unwrap {'text': '...'} wrapper	—
2	NFC normalize	Unicode NFC + control chars + whitespace normalization	—
3	Min length	Minimum text length and word count	≥50 chars, ≥10 words
4	Kazakh chars	Must contain ≥1 Kazakh-specific character (Ә, Ғ, Қ, Ң, Ө, Ұ, Ү, Һ, І)	≥1 char
5	Script profile	Cyrillic ≥60%, Latin ≤25%	cyr≥0.60, lat≤0.25
6	Junk filter	URL density, HTML tags, special char ratio, boilerplate	URL≤5/1K, HTML≤5, special≤40%
7	Gzip repetition	Compression ratio to detect repetitive text	ratio≥0.20
8	FastText LID	Language identification (kk≥0.5, gap to rivals ≥0.1)	kk≥0.50, gap≥0.10
9	Exact dedup	MD5 hash deduplication across all sources	—

Long texts (>50K chars) are split into chunks at paragraph/sentence boundaries before filtering.

Top rejection reasons

Reason	Count	% of rejected
no_kaz_chars	7,713,488	52.3%
dedup	3,170,330	21.5%
too_few_words	1,672,400	11.3%
too_short	1,627,752	11.0%
lid_rejected	246,943	1.7%
script_profile	195,774	1.3%
gzip_repetition	127,404	0.9%
junk	114,098	0.8%

Schema

{
    "text": str,    # Cleaned text
    "source": str   # Source identifier (e.g., "culturax", "md_oscar", "cc100")
}

Usage

from datasets import load_dataset

ds = load_dataset("stukenov/sozkz-corpus-clean-v3")

train = ds["train"]
val = ds["validation"]

# Check source distribution
from collections import Counter
counts = Counter(train["source"])
for src, n in counts.most_common():
    print(f"{src:20s}: {n:>10,d}")

# Sample texts
for i in range(3):
    sample = train[i]
    print(f"[{sample['source']}] {sample['text'][:200]}")

Intended Use

Pre-training small-to-medium language models on Kazakh text. The corpus provides broad coverage across web text, news, literature, encyclopedic content, and parallel corpora.

Limitations

• No quality scoring: All texts that pass filters are treated equally; no quality-based weighting
• Domain imbalance: Web crawl sources (culturax, madlad, hplt) dominate the corpus
• LID errors: FastText LID may misclassify some Kazakh texts as related Turkic languages (Kyrgyz, Bashkir)
• Parallel corpus residue: kazparc/kazparc_sync texts are Kazakh-side extracts from parallel data; some may lack natural flow

Citation

@dataset{sozkz_corpus_clean_v3_2026,
  title={SozKZ Corpus Clean v3: Cleaned Kazakh Text Corpus},
  author={Saken Tukenov},
  year={2026},
  url={https://huggingface.co/datasets/stukenov/sozkz-corpus-clean-v3}
}