sozkz-corpus-dedup-kk-web-v1

Deduplicated Kazakh web text corpus collected from 6 public HuggingFace datasets. Contains only texts not present in kz-transformers/multidomain-kazakh-dataset (12.4M texts were used as the dedup reference).

Stats

Field	Value
Total unique texts	9,475,089
Format	Parquet (142 shards)
Columns	text, source
Dedup method	MD5 hash (exact match)
Dedup reference	kz-transformers/multidomain-kazakh-dataset (12.4M hashes)
Date	2026-02-13
Version	v1
License	Apache 2.0

Sources

Source	HF Dataset	Unique texts	Dupes removed
culturax	uonlp/CulturaX (kk)	2,705,991	25,943
hplt	HPLT/HPLT2.0_cleaned (kaz_Cyrl)	2,246,264	391,066
mc4	allenai/c4 (kk)	2,230,795	140,733
madlad400	allenai/MADLAD-400 (kk)	1,807,827	169
moscar	oscar-corpus/mOSCAR (kaz_Cyrl)	245,869	0
wikipedia	wikimedia/wikipedia (20231101.kk)	238,343	13

Cleaning

• Unicode NFC normalization
• Control character removal
• Whitespace/newline collapsing
• Min text length: 50 chars
• Max URL density filter
• HTML tag filter

Usage

from datasets import load_dataset

ds = load_dataset("stukenov/sozkz-corpus-dedup-kk-web-v1", split="train")
print(len(ds))  # 9,475,089
print(ds[0]["text"][:200])
print(ds[0]["source"])  # e.g. "culturax"

Complementary dataset

This corpus is designed to complement kz-transformers/multidomain-kazakh-dataset. Together they provide ~21.9M unique Kazakh texts.

Related repos

• Tokenizer: stukenov/kazakh-bpe-32k
• Base model: stukenov/sozkz-core-llama-50m-kk-base-v1