Datasets:

metinovadilet
/

duplicate-of-kyrgyz-ner

Modalities:

Text

Formats:

text

Size:

100K - 1M

ArXiv:

Libraries:

Datasets

Dataset card Data Studio Files Files and versions

xet

Community

Dataset Viewer

Auto-converted to Parquet Duplicate

Split (2)

train · 103k rows

text stringlengths 0 42
# 1 0 1
Жалал - - B-LOCATION
- - - I-LOCATION
Абад - - I-LOCATION
облусундагы - - I-LOCATION
Ала - - I-LOCATION
- - - I-LOCATION
Бука - - I-LOCATION
— - - I-LOCATION
Жаңы - - I-LOCATION
- - - I-LOCATION
Базар - - I-LOCATION
— - - I-LOCATION
Кировка - - I-LOCATION
жолунда - - I-LOCATION
тоодон - - O
кулаган - - O
таш - - O
аралаш - - O
кум - - O
- - - O
шагыл - - O
бир - - B-MEASURE
жүк - - O
ташуучу - - O
автоунааны - - O
басып - - O
калды - - O
. - - O

# 2 0 2
Кырсык - - O
16-майда - - B-PERIOD
түндө - - I-PERIOD
жолдун - - O
Чаткал - - B-LOCATION
районундагы - - I-LOCATION
Терек - - I-LOCATION
- - - I-LOCATION
Сай - - I-LOCATION
айыл - - I-LOCATION
аймагында - - O
болгон - - O
. - - O

# 3 0 3
Өзгөчө - - B-INSTITUTION
кырдаалдар - - I-INSTITUTION
министрлигинин - - I-INSTITUTION
басма сөз - - O
кызматы - - O
билдиргендей - - O
, - - O
машинеде - - O
30 - - B-MEASURE
жаштагы - - O
айдоочудан - - O
тышкары - - O
дагы - - O
бир - - B-MEASURE
адам - - O
бар болчу - - O
. - - O

# 4 0 4
Экөөнүн - - B-MEASURE
тең - - O
абалы - - O
дурус - - O
. - - O

# 5 0 5
Учурда - - O
жол - - O
тазаланууда - - O
. - - O

# 6 0 6
Бирок - - O
таш - - O
кулоо - - O
коркунучу - - O
бар болгонуна - - O
байланыштуу - - O
тазалоо - - O
иштерине - - O
тоскоолдук - - O
жаралып - - O
атат - - O
. - - O

# 7 1 1
2,5 - - B-MEASURE
миңдей - - I-MEASURE
кыргызстандыктар - - B-PERSON_TYPE
диний - - O
ишениминен - - O
улам - - O
жаш - - O
балдарын - - O

End of preview. Expand in Data Studio

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Kyrgyz NER

Task

Named Entity Recognition (Token Classification)

Description

The first manually annotated NER dataset for the Kyrgyz language. Contains 1,499 Kyrgyz news articles from the 24.kg news portal (2017-2022), annotated by 59 trained Kyrgyz linguists with an inter-annotator agreement of κ = 0.89.

27 entity types: Person, Location, Institution, Measure, Period, Event, Law, Product, etc.
Format: CoNLL-2003 (token-per-line, tab-separated)
Native Kyrgyz text (not translated)

Dataset Size

Split	Docs	Sentences	Tokens	Entity Mentions
Train	999	7,033	89,248	24,949
Test	500	3,867	51,118	14,126

Base Model

KyrgyzBERT (35.9M params)

HuggingFace: metinovadilet/KyrgyzBert
Tokenizer: metinovadilet/bert-kyrgyz-tokenizer

Benchmark Against

Model	Params	Type	Published F1
KyrgyzBERT	35.9M	Monolingual Kyrgyz	— (new)
mBERT (bert-base-multilingual-cased)	178M	Multilingual	0.68
XLM-RoBERTa-base	278M	Multilingual	0.73
CRF	—	Traditional	0.62

Published baselines are from the KyrgyzNER paper (Turatali et al., 2025).

Goal

Demonstrate that KyrgyzBERT achieves competitive fine-grained NER performance on native Kyrgyz text against models 5-8x its size. A strong result here would be especially significant since this is human-annotated native Kyrgyz data, not translated.

License

Dataset: CC BY-NC-SA 4.0
Code & Models: MIT

Citation

@inproceedings{turatali2025kyrgyzner,
  title     = {Human-Annotated NER Dataset for the Kyrgyz Language},
  author    = {Turatali, Timur and Alekseev, Anton and Jumalieva, Gulira and Kabaeva, Gulnara and Nikolenko, Sergey},
  booktitle = {Proceedings of TurkLang 2025},
  year      = {2025}
}

Source

GitHub: https://github.com/Akyl-AI/KyrgyzNER
Paper: https://arxiv.org/abs/2509.19109
HuggingFace Model: https://huggingface.co/TTimur/xlm-roberta-base-kyrgyzNER

Downloads last month: 99

Paper for metinovadilet/duplicate-of-kyrgyz-ner

Human-Annotated NER Dataset for the Kyrgyz Language

Paper • 2509.19109 • Published Sep 23, 2025