Til-Parallel-KK — Kazakh ↔ Russian ↔ English parallel corpus

Sentence pairs for machine translation involving Kazakh, in four directions: Kazakh to and from Russian, and Kazakh to and from English. Each pair carries its translation direction, a subject domain and a quality score.

Sentences are short — median 105 characters, roughly one sentence per pair.

Quick start

from datasets import load_dataset

ds = load_dataset("TilQazyna/Til-Parallel-KK", "clean", split="train")

# one direction only
ru_kk = ds.filter(lambda r: r["task"] == "translate_ru_kk")
print(ru_kk[0])
# {'input':  'Использование текущих и постоянных цен помогает нам прояснить это различие.',
#  'target': 'Ағымдағы және тұрақты бағаларды пайдалану бұл айырмашылықты нақтылауға көмектеседі.',
#  'task': 'translate_ru_kk', 'source': 'gec-mix', 'split': 'train',
#  'score': 4, 'category': 'economy_finance', 'judge_lang': 'kk'}

Configurations

Three quality tiers of the same collection, nested rather than disjoint.

Config	Rows	Use it when
clean (default)	664 555	General training
premium	388 552	Highest-confidence subset only
raw	707 773	You want to filter it yourself

Translation directions

Direction	raw	clean	premium
translate_en_kk	180 148	170 840	115 233
translate_kk_en	179 983	157 186	64 657
translate_ru_kk	174 233	169 281	115 590
translate_kk_ru	173 409	167 248	93 072

The four directions are close to balanced in raw. In premium the Kazakh→English side thins out considerably, so if you are training that direction specifically, prefer clean.

Data fields

Field	Type	Description
input	string	Source sentence
target	string	Translated sentence
task	string	Direction: translate_ru_kk, translate_kk_ru, translate_en_kk, translate_kk_en
score	int64	Quality rating 0–5 assigned during curation
category	string	Subject domain, e.g. everyday_life, literature
judge_lang	string	Language detected in the target side
source	string	Always gec-mix, an artefact of the collection's history (see below)
split	string	Always train
error_tags	string	Always null. Left over from the schema this data used to live in; ignore it.

The source and error_tags fields are kept only so that rows can still be traced back to the predecessor dataset. Neither carries information here.

Domains and quality

Domains lean towards everyday and cultural text rather than officialese: everyday_life (202 345), literature (173 862), medicine_health (167 447), history (139 237), technology_it (117 234), economy_finance (95 420), entertainment (87 866), science_academic (81 679).

Scores: 5 → 654 570, 4 → 511 086, 3 → 552 006, and 43 218 pairs at 0–2. Filtering on score >= 4 keeps roughly two thirds of the corpus.

How this dataset came to exist

These pairs were published inside TilQazyna/Til-GEC, a grammatical error correction dataset, where they made up about 39% of the rows. They had ended up there because the collection that supplied them (stukenov/sozkz-corpus-pretrain-gec-mix-v1) was a pretraining mixture rather than a correction set, and was merged in on the strength of its name.

The rows were always labelled correctly — every one of them carried a translate_* value in task — so extracting them was an exact filter on that field, not a guess. They are published here as their own dataset because a parallel corpus of this size for Kazakh is worth having on its own terms, and because leaving it inside a GEC dataset actively harmed models trained there.

The correction half is published as TilQazyna/Til-GEC-v2.

Limitations

• No held-out split. Everything is train. Carve out your own evaluation set, and check for overlap if you also evaluate on FLORES, NTREX or similar — provenance of the underlying text is not fully documented.
• Scores are machine-assigned, not human-verified, and quality within a tier varies.
• Direction is a label, not a guarantee. Spot-checking is advisable before training a production system on any single direction.
• English↔Kazakh pairs may be pivoted through a third language rather than translated directly. This is not recorded per row.
• Single reference per source sentence.

Related datasets

Dataset	What it is
TilQazyna/Til-GEC-v2	The grammatical-correction half of the same original collection
TilQazyna/Til-GEC	The collection both were split out of, deprecated

Citation

@misc{tilparallelkk_2026,
  title  = {Til-Parallel-KK: A Kazakh--Russian--English Parallel Corpus},
  author = {TilQazyna},
  year   = {2026},
  url    = {https://huggingface.co/datasets/TilQazyna/Til-Parallel-KK}
}

---

По-русски, кратко

Параллельный корпус для машинного перевода с казахским в четырёх направлениях: kk↔ru и kk↔en. У каждой пары есть направление, домен и оценка качества. Предложения короткие — медиана 105 символов.

Три конфигурации: clean (664 555 пар, берите по умолчанию), premium (388 552, строгий отбор), raw (707 773, всё). В raw направления почти сбалансированы, а в premium сильно проседает kk→en — для этого направления лучше брать clean.

Корпус выделен из TilQazyna/Til-GEC, где он лежал под видом данных для исправления грамматических ошибок и занимал там около 39% строк. Метка направления в поле task была проставлена у всех строк, поэтому выделение — точный фильтр, а не догадка.

О чём стоит знать заранее: отложенной выборки нет, всё в train — свою придётся выделять самим. Оценки проставлены машиной, а не людьми. Пары en↔kk могли получиться через третий язык, и в данных это не отмечено.