You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Log in or Sign Up to review the conditions and access this dataset content.
EkiTil Parallel Corpus kk-ru/kk-en v2
A large-scale deduplicated parallel corpus for Kazakh-Russian and Kazakh-English machine translation.
Dataset Summary
| Config |
Pairs |
Languages |
| kk-ru |
5,100,973 |
Kazakh ↔ Russian |
| kk-en |
126,149 |
Kazakh ↔ English |
Sources (kk-ru)
| Source |
Pairs |
Description |
| WMT19 crawl |
4,512,841 |
Web-crawled kk-ru parallel sentences (statmt.org) |
| KazParC |
362,208 |
Human-translated, 5 domains (issai/kazparc) |
| OPUS-XLEnt |
75,837 |
Cross-lingual entity pairs |
| OPUS-KDE4 |
53,912 |
KDE software localization |
| OPUS-wikimedia |
43,173 |
Wikimedia content |
| OPUS-WikiMatrix |
32,786 |
Mined Wikipedia parallel sentences |
| OPUS-TED2020 |
5,880 |
TED talk subtitles |
| OPUS-GNOME |
3,514 |
GNOME software localization |
| OPUS-QED |
3,488 |
Educational video subtitles |
| OPUS-NeuLab |
2,798 |
TED talks (NeuLab) |
| OPUS-Tatoeba |
2,310 |
Community-contributed translations |
| OPUS-OpenSubtitles |
2,089 |
Movie/TV subtitles |
| OPUS-Ubuntu |
137 |
Ubuntu software localization |
Sources (kk-en)
| Source |
Pairs |
Description |
| WMT19 |
126,149 |
WMT19 shared task kk-en parallel data |
Preprocessing
- Cleaning: removed control characters, normalized whitespace, filtered pairs where source = target
- Length filtering: removed pairs shorter than 3 chars or longer than 5000 chars per side
- Ratio filtering: removed pairs where one side is >10x longer than the other
- Deduplication: exact-match dedup by MD5 hash of (kk, ru) pair
- Shuffling: randomly shuffled with seed=42
Usage
from datasets import load_dataset
ds = load_dataset("stukenov/ekitil-parallel-kkru-v2", "kk-ru", split="train")
print(ds[0])
ds_en = load_dataset("stukenov/ekitil-parallel-kkru-v2", "kk-en", split="train")
Columns
| Column |
Type |
Description |
| kk |
string |
Kazakh text |
| ru / en |
string |
Russian or English text |
| source |
string |
Data source identifier |
| domain |
string |
Domain category |
Domain Distribution (kk-ru)
- web (88.5%): government websites, legal documents, news — WMT19 crawl
- legal_docs (2.9%): legal and regulatory texts — KazParC
- mass_media (2.4%): news articles — KazParC
- general (4.4%): mixed OPUS corpora
- edu_and_sci (0.9%): educational content — KazParC
- fiction (0.6%): literary works — KazParC
Intended Use
- Training machine translation models (kk↔ru, kk↔en)
- Fine-tuning bilingual language models for translation tasks
- Research in low-resource machine translation
Citation
@misc{ekitil-parallel-2026,
title={EkiTil Parallel Corpus for Kazakh-Russian Machine Translation},
author={Saken Tukenov},
year={2026},
url={https://huggingface.co/datasets/stukenov/ekitil-parallel-kkru-v2}
}
License
MIT
