Til-GEC-v2 — Kazakh grammatical error correction

Sentence pairs for correcting grammar and spelling in Kazakh: an input sentence that contains an error and a target sentence that fixes it. Each pair is labelled with the type of error and the subject domain it came from.

About 14% of the pairs (401 661) are identity pairs — input and target are the same sentence, labelled identity. They are there on purpose: a corrector that cannot leave a correct sentence alone is useless, and a model trained without them rewrites everything.

Quick start

from datasets import load_dataset

ds = load_dataset("TilQazyna/Til-GEC-v2", "clean", split="train")
print(ds[0])
# {'input':  'Өзен сағасы Вига өзенынің сол жағалауынан 66 км қашықтықта орналасқан.',
#  'target': 'Өзен сағасы Вига өзенінің сол жағалауынан 66 км қашықтықта орналасқан.',
#  'error_tags': '["orthography/vowel_harmony"]',
#  'task': 'gec', 'source': 'gec-rulebased', 'split': 'train',
#  'score': 5, 'category': 'geography_travel', 'judge_lang': 'kk'}

# The fix above is a single letter: өзенынің -> өзенінің. Kazakh suffixes must agree
# with the vowels of the stem, and that is the single most common real error class here.

Configurations

Three quality tiers of the same collection. They are nested, not disjoint — premium is the strictest selection, raw is everything.

Config	Rows	Use it when
clean (default)	1 018 673	General training. Deduplicated, filtered.
premium	758 974	You want the highest-confidence subset only.
raw	1 082 457	You want to apply your own filters from scratch.

Data fields

Field	Type	Description
input	string	Kazakh sentence, possibly containing an error
target	string	The corrected sentence
error_tags	string	A JSON-encoded array of error labels, e.g. '["morphosyntax/case"]'. Parse it with json.loads.
task	string	Always gec in this dataset
source	string	Which sub-collection the pair came from (see below)
split	string	Always train — no official test split ships here, see Evaluation
score	int64	Quality rating 0–5 assigned during curation
category	string	Subject domain, e.g. politics, medicine_health
judge_lang	string	Language detected in the pair, kk for almost all rows

What is in the data

Error types (top labels, clean + premium + raw combined):

Label	Pairs
orthography/spelling	470 122
identity (no error — leave unchanged)	401 661
morphosyntax/case	335 021
orthography/vowel_harmony	261 253
orthography/spacing	226 561
orthography/punctuation	173 581
syntax_discourse/word_order	111 674
morphosyntax/plural	90 445
syntax_discourse/missing_element	88 386
syntax_discourse/redundant_element	66 115
morphosyntax/agreement	53 224
morphosyntax/possessive	37 402
morphosyntax/tense	28 731
morphosyntax/postposition	19 001

orthography/vowel_harmony is specific to Kazakh: suffixes must agree with the vowels of the stem, and getting that wrong is one of the most common real mistakes in written Kazakh.

Domains — news and public-sector text dominate: politics (367 109), economy_finance (240 215), education (216 839), news_media (203 412), medicine_health (202 092), geography_travel (188 053), everyday_life (148 236), government (137 481), law (134 892).

Quality scores — most pairs sit at the top of the scale: 5 → 1 498 308, 4 → 778 614, 3 → 519 398, and 63 784 pairs at 0–2. Filter on score >= 4 if you want the stricter half without switching configs.

Sub-collections (source field): gec-rulebased 2 807 468, gec-synth 52 630, gec-gpt4o 6.

How this dataset was built

This is a filtered version of TilQazyna/Til-GEC. Two groups of rows were removed from the predecessor:

1. 707 773 translation pairs. The predecessor was assembled from several sources, and one of them (sozkz-corpus-pretrain-gec-mix-v1) was a pretraining mixture, not a correction set. Those rows carried task = translate_ru_kk / translate_kk_ru / translate_en_kk / translate_kk_en and had no error labels at all. Training on them teaches a model to translate rather than to correct. They were not discarded — they are published separately as TilQazyna/Til-Parallel-KK.
2. 1 436 benchmark rows. The predecessor contained rows from the public evaluation set stukenov/sozkz-corpus-gec-benchmark-kk-v1. An evaluation set inside a training set silently inflates every score measured against it.

Rows were routed by the task field, which was intact in the source data, so the split is exact rather than heuristic. Nothing else was modified: the remaining rows, their fields and their values are byte-identical to the predecessor.

Evaluation

There is no test split here on purpose. Evaluate against stukenov/sozkz-corpus-gec-benchmark-kk-v1, which has test, organic_social (real mistakes collected from social media) and synthetic splits. Report organic_social alongside test — the two behave very differently, and a model that looks strong on synthetic errors can be weak on real ones.

Limitations

• The errors are mostly generated, not observed. gec-rulebased accounts for 98% of the data: real Kazakh sentences with errors introduced by rule. The distribution of these errors does not match how people actually write, and models trained on this data tend to do markedly worse on genuine human mistakes than on synthetic ones.
• Domain skew. News, politics and public-sector text are heavily over-represented. Conversational Kazakh is thin.
• Single reference. Each input has exactly one accepted correction, though several may be valid. Exact-match scoring will therefore understate real performance.
• Scores are machine-assigned, not human-verified.

Related datasets

Dataset	What it is
TilQazyna/Til-Parallel-KK	The translation pairs removed from this collection, kk↔ru and kk↔en
stukenov/sozkz-corpus-gec-benchmark-kk-v1	Evaluation benchmark, excluded from this dataset
TilQazyna/Til-GEC	Predecessor, deprecated

Citation

@misc{tilgec_v2_2026,
  title  = {Til-GEC-v2: A Kazakh Grammatical Error Correction Corpus},
  author = {TilQazyna},
  year   = {2026},
  url    = {https://huggingface.co/datasets/TilQazyna/Til-GEC-v2}
}

---

По-русски, кратко

Пары предложений для исправления грамматики и орфографии в казахском языке: на входе предложение с ошибкой, на выходе — исправленное, плюс метка типа ошибки и домен. Около 14% пар (401 661) — это пары без ошибки, где вход и выход совпадают: корректор, который не умеет оставлять правильный текст в покое, бесполезен и правит всё подряд.

Три конфигурации — тиры качества одной и той же коллекции: clean (1 018 673 строки, берите её по умолчанию), premium (758 974, самый строгий отбор), raw (1 082 457, всё сырьё).

Это очищенная версия TilQazyna/Til-GEC. Из предшественника убрано 707 773 переводные пары, попавшие туда из претрейн-микса sozkz-corpus-pretrain-gec-mix-v1 — они вынесены в отдельный датасет TilQazyna/Til-Parallel-KK, — и 1 436 строк публичного бенчмарка, которым эту же задачу и меряют.

О чём стоит знать заранее: 98% ошибок здесь порождены правилами, а не собраны у живых людей. Их распределение не совпадает с тем, как люди ошибаются на самом деле, поэтому модель, обученная только на этих данных, обычно заметно слабее на настоящих ошибках, чем на синтетических. Меряйте на сплите organic_social бенчмарка, а не только на test.