Datasets:

podarok
/

kobza-cleaned-ua

license: cc-by-4.0
language:
  - uk
size_categories:
  - 1M<n<10M
task_categories:
  - text-generation
tags:
  - ukrainian
  - web-crawl
  - cleaned
  - language-modeling
dataset_info:
  features:
    - name: text
      dtype: string
    - name: source
      dtype: string
    - name: length
      dtype: int64
  splits:
    - name: train
      num_bytes: 536853186
      num_examples: 1812460
  download_size: 248201329
  dataset_size: 536853186
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

kobza-cleaned-ua

Cleaned Ukrainian-language subset of Goader/kobza dataset with Russian content filtered out.

Dataset Details

Dataset Description

This dataset is a cleaned and filtered version of the Goader/kobza corpus, removing Russian language content to create a pure Ukrainian language dataset suitable for training language models.

The original kobza dataset contains ~60B tokens across 97 million documents. This cleaned version maintains ~59B tokens (98.5% retention) after removing Russian content detected through character patterns, word lists, and grammatical structures.

Curated by: podarok
Language(s): Ukrainian (uk)
License: CC-BY-4.0

Dataset Sources

Repository: https://github.com/podarok/ua_ai_v2
Parent Dataset: Goader/kobza

Uses

Direct Use

This dataset is intended for:

Pre-training Ukrainian language models
Fine-tuning multilingual models on Ukrainian text
Ukrainian NLP research and development
Training text generation models

Out-of-Scope Use

This dataset should not be used for:

Applications requiring guaranteed absence of any Russian text (some edge cases may remain)
Applications requiring quality scoring (no quality scores provided)
Real-time applications (streaming recommended for large-scale use)

Dataset Structure

The dataset contains 1,812,460 documents with the following schema:

{
    "text": str,      # Document text
    "source": str,    # One of: hplt-2.0, fineweb-2, cultura-x, ubertext2.0, ukrainian-news
    "length": int     # Character count
}

Source Distribution

Source	Documents	Size (MB)	Russian Removed
hplt-2.0	641,685	160.6	1.77%
fineweb-2	539,769	154.0	1.77%
cultura-x	383,117	119.6	2.23%
ubertext2.0	192,010	21.6	0.20%
ukrainian-news	55,879	14.8	0.88%
Total	1,812,460	470.6	~1.5%

Dataset Creation

Curation Rationale

The original Goader/kobza dataset, while being the largest Ukrainian corpus, contains some Russian language content due to:

Mixed-language websites
Code-switching in web content
Multilingual web crawls

This cleaned version was created to provide a higher-quality, Ukrainian-only corpus for language model training.

Source Data

Data Collection and Processing

The source data comes from the Goader/kobza dataset, which aggregates text from:

hplt-2.0: High-quality web crawl data
fineweb-2: Curated web content
cultura-x: Cultural and literary texts
ubertext2.0: Ukrainian language corpus
ukrainian-news: News articles

Filtering Process:

Russian-only characters: Detection of Cyrillic characters exclusive to Russian (ы, э, ё, Э, Ы, Ё)
Russian word patterns: Removal of common Russian-only words (и, что, да, можно, etc.)
Mixed-language detection: Filtering sentences with Russian grammatical patterns
Line-by-line filtering: Each line evaluated independently

Who are the source data producers?

The source data was originally curated by Goader from various web sources. The cleaning and filtering was performed by the ua_ai_v2 project team.

Annotations

This dataset does not contain annotations beyond the source metadata.

Personal and Sensitive Information

The dataset inherits any personal or sensitive information present in the original Goader/kobza dataset. Users should refer to the original dataset's documentation for details on privacy considerations.

Bias, Risks, and Limitations

Incomplete filtering: Some Russian text may remain due to edge cases or similar words between languages
No quality scores: Unlike some corpora, this dataset does not include document-level quality scores
Source bias: Inherits any biases present in the original web crawl sources
Temporal bias: Reflects web content from the time period of the original crawl
Domain distribution: Web content is heavily represented compared to other text types

Recommendations

Users should:

Validate the dataset's suitability for their specific use case
Consider combining with other Ukrainian corpora for better domain coverage
Apply additional quality filtering if needed for production use
Be aware that some Russian content may remain in edge cases

Usage

Basic Loading

from datasets import load_dataset

dataset = load_dataset("podarok/kobza-cleaned-ua", split="train")
print(f"Total documents: {len(dataset):,}")

Streaming Mode (Recommended for Large-Scale Training)

from datasets import load_dataset

dataset = load_dataset("podarok/kobza-cleaned-ua", split="train", streaming=True)
for doc in dataset:
    print(doc["text"])

Filter by Source

# Get only news content
news_dataset = dataset.filter(lambda x: x["source"] == "ukrainian-news")

# Get multiple sources
web_dataset = dataset.filter(
    lambda x: x["source"] in ["hplt-2.0", "fineweb-2"]
)

Citation

Original Dataset

@misc{kobza2024,
  title={Kobza: Ukrainian Language Corpus},
  author={Goader},
  year={2024},
  url={https://huggingface.co/datasets/Goader/kobza}
}

This Dataset

If you use this cleaned version, please cite both the original kobza dataset and reference this filtered version.

Dataset Card Authors

podarok (cleaning and curation)

Dataset Card Contact

For questions or issues, please open an issue at https://github.com/podarok/ua_ai_v2