Datasets:

SoyVitou
/

KEPTA

id int64 1 3.38k	label stringlengths 1 123	language class label 3 classes
781	យោង :	0Khmer
78	Thus the wise win before	2English
3,251	១៨៥០ រហូតមកទល់សព្វថ្ងៃ ។	0Khmer
1,889	10	2English
789	ទីក្រុងគូឡាឡាំពួ -គេហទំព័រ និងបណ្ដាញសង្គម៖	0Khmer
1,589	79.7	2English
1,021	ជាភពដែលជិតព្រះអាទិត្យជាងគេ និងគ្មាន	0Khmer
2,885	បទប្បញ្ញត្តិនោះ។	0Khmer
71	do not become angered	2English
79	the fight, while	2English
1,703	mode rather than one-way output and input mode.	2English
2,755	និងកំណត់ពីការផ្តន្ទា	0Khmer
2,428	2024-12-08.	2English
1,172	degrees than both projections will return a mess. In that case you can cut your document to	2English
2,957	ផង។	0Khmer
1,900	២០១៤	0Khmer
51	សំណួរ	0Khmer
2,252	AMC Benchmark: The surest way to guard against overfitting to the test set is to test on fresh data.	2English
353	via layers of residual convolutional blocks and a layer of gated recurrent units	2English
243	អេស៊ីអ៊ី ទួលទំពួង	0Khmer
3,265	វណ្ណៈអ្នកដឹកនាំនិងបង្កើតទុរគតភាពនៃរណ្ណៈពលករ ។ ទាំងនេះជាប្រភពនាំឱ្យមានចលនា	0Khmer
1,285	method, CNN predicts the angle of landscape photos without any preprocessing. It is also clear	2English
984	ច្បាប់ស្ដីពីការបង្កើតក្រសួងអប់រំ យុវជន និងកីឡា	0Khmer
420	facilitate the teaching and learning of the Khmer language among a group of Grade 1 students	2English
2,900	មាត្រា ២០- ជនរងគ្រោះជាជនជាតិខ្មែរ	0Khmer
592	សមាគមអ្នកអក្សរសិល្ប៍វៀតណាម ឆ្នាំ២០១៩។	0Khmer
130	worry and think before making	2English
3,068	ពុំមានបញ្ហាអស្តិរូបទេ ដែលគេចាត់បញ្ចូលទៅក្នុងទស្សនវិជ្ជារូបនិយម។ អាស្រ័យហេតុនេះ	0Khmer
2,815	cc.feedback2011 @gmail.com។	1Khmer + English
3,376	បណ្ដុះទស្សនគតិដើម្បីភាពជោគជ័យ	0Khmer
1,386	5. SIMULATION RESULTS	2English
2,390	Token Search (PTS), to generate pairs for DPO for the first DPO round. Details of the data mixture	2English
423	for the Grade 1 to 9 Khmer textbooks showed that the frequency of word exposure and the	2English
1,917	៥,០	0Khmer
2,739	ជាចុងក្រោយ ខ្ញុំក៏សូមថ្លែងអំណរគុណដ៏ជ្រាលជ្រៅដល់	0Khmer
410	regard to the usability and learnability of the Khmer language by focusing on vocabulary size,	2English
1,412	7	2English
193	សូមគោរពជូន	0Khmer
2,679	ក្រមព្រហ្មទណ្ឌ	0Khmer
2,223	ទាញចេញពីលទ្ធផលនៃការសិក្សាស្រាវជ្រាវនៅក្នុងខ្លឹមសារខាងលើ យើងខ្ញុំមានអនុសាសន៍	0Khmer
2,129	ការសិក្សា នៅមធ្យមសិក្សា មនុស្សមួយ	0Khmer
1,406	6. Images with -90-degrees orientation predicted in 90-degrees category.	2English
1,158	Khmer (ភាសាខ្មែរ):	1Khmer + English
3,062	ចែកអក្សរសិល្ប៍និងលក្ខណតថៈរបស់រឿង ចលនាអក្សរសិល្ប៍ ប្រធានរឿង កាលអាកាស មូល	0Khmer
1,815	5,000	2English
9	ExtraLight ប្រទេសកម្ពុជា ១២៣ Cambodia 123	1Khmer + English
622	[WFM+24] Yue Wu, Yewen Fan, So Yeon Min, Shrimai Prabhumoye, Stephen McAleer, Yonatan Bisk,	2English
1,341	and therefore shows no impact towards classification score (zero correlated) or may have a	2English
3,266	អក្សរសាស្ត្រថ្មី ។	0Khmer
1,391	sad().stop();	2English
2,543	Generation 8, Class E4, Information Technology of Engineering Department	2English
2,193	បើកទូលាយឱ្យមានការចូលរួមពីក្រសួងស្ថាប័នគ្រប់ដៃគូអភិវឌ្ឍន៏និងសង្គមស៊ីវិល	0Khmer
1,696	by strengthening bilateral cooperative ties. In addition, technical training and cultural	2English
1,310	and completely utilize the data of the content line.	2English
2,258	of original prompts written by the team (see Section 5 for further details).	2English
2,054	និង	0Khmer
126	A successful samurai should put	2English
552	hakaran, V., Reif, E., Du, N., Hutchinson, B., Pope, R., Bradbury, J., Austin,	2English
2,613	meet market demands, (2) increase equitable access to higher education, and (3) strengthen the	2English
2,777	ថាច្បាប់នេះត្រូវបានរដ្ឋសភាអនុម័តជាភាសាខ្មែរ នាំឲ្យអត្ថបទ	0Khmer
3,132	ដំណើរតាមប្តឹមកស្រុកចាន់តុង" ។	0Khmer
2,738	មួយចំនួនក្នុងអម្លងពេលនៃការបកប្រែក្រមនេះ។	0Khmer
842	៥. តើអ្នកចូលចិត្តពុម្ពអក្សរមួយណា? សូមប្រាប់ឲ្យយើងដឹងផង។	0Khmer
2,683	រាជរដ្ឋាភិបាល	0Khmer
2,544	Faculty of Engineering, Royal University of Phnom Penh	2English
1,771	ឧស្សាហកម្ម	0Khmer
3,324	សពលភាព: ១៣.០៩.២០១៥ ដល់ថ្ងៃ ១២.០៩.២០២៥	0Khmer
477	ដែលអាចផ្ទៀងផ្ទាត់បានតាមរយៈ "ថ្នាលផ្ទៀងផ្ទាត់ឯកសារ verify.gov.kh"	1Khmer + English
1,918	៧,១	0Khmer
3,275	១	0Khmer
2,199	១៦.១	0Khmer
436	Numerous highly stressful events occurred in Cambodia in the last century.	2English
2,596	• The manufacturer of the offered goods, if a different entity from the Bidder, has	2English
2,466	model. In Thirty-seventh Conference on Neural Information Processing Systems, 2023.	2English
1,113	\section{Strengths and Limitations of the OCR System}	2English
176	The test environment is configured on	2English
1,708	Considering the organization and action capacity of the Lao government, it is	2English
29	ពុម្ពអក្សរនគរ	0Khmer
18	កម្ពុជាត្រូវបានគេស្គាល់ជាទូទៅថា Cambodia ជាភាសាអង់គ្លេស	1Khmer + English
1,723	between the two sides should be two-way interaction, not one-way. In the process of	2English
2,562	Copy	2English
1,986	រៀបចំរចនាសម្ព័ន្ធគាំទ្រដល់កម្លាំងពលកម្ម	0Khmer
620	Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, et al.	2English
784	មន្ត្រី រដ្ឋមន្ត្រីក្រសួង អប់រំ យុវជននិងកីឡា បានអញ្ជើញអម	0Khmer
610	The phi-4 model is based on a decoder-only transformer architecture [VSP+17] with 14B parameters and	2English
1,283	identifying the skew angles using various computer vision algorithms. Fischer et al. [1] applied	2English
3,003	ចលនារាង	0Khmer
1,530	[2] Ramesh, D., Sanampudi, S.K.: An automated essay scoring systems: a systematic	2English
1,119	print(df_train.head())	2English
1,368	CLOUD	2English
1,491	processing, and the goal of increased computational power. However, the answer could	2English
467	texts, as shown below.	2English
3,110	សំពត់សូត្រ យកទៅដូរឱ្យអ្នករស់នៅតាមភូមិភាគខាងជើង។ ពួកគេដឹកនាំយកមកវិញនូវភោគទ្រព្យ	0Khmer
1,376	Interpretations	2English
2,440	arXiv:2009.03300, 2020.	2English
294	ទីតាំងប្រើប្រាស់ចរន្ត	0Khmer
773	ថ្នាក់មេត្តាជ្រាប់ថា ៖ ក្រុមហ៊ុន QUAN DAT នឹងធ្វើការសម្អាតជញ្ជាំងកញ្ចក់ និងអាលុយកាបោន អគារថ្មី	1Khmer + English
2,283	requires curating diverse seeds from organic sources.	2English
1,156	manager. He rubbed his hands all the time as if	2English
1,273	file_panel_item_rename = ['ctrl+r', '']	2English

End of preview. Expand in Data Studio

Khmer OCR Dataset Collection (KEPTA)

KEPTA stands for Khmer-English Printed Text Annotation.

This dataset and repository provide Khmer OCR annotation assets and utilities for preparing, cleaning, visualizing, and uploading the dataset to Hugging Face.

Why this dataset exists

Khmer is a low-resource language, and publicly available OCR benchmarks are limited.
This dataset is prepared to serve as a benchmark to enable fair comparison of OCR models and to support research and development on Khmer text recognition.
Benchmark preparation by: Soy Vitou and Eong Koungmeng.

Dataset summary

Task: OCR / text recognition (printed text crops)
Columns: id, image, label, language
Language labels: Khmer; Khmer + English; English

Dataset layout

annotation/label.txt: Tab-separated image_path<TAB>label; built from images.db plus the Gemini dataset (primary first). Crops live under annotation/image/....
annotation/normalization.txt: Normalized labels produced by normalization.py.
annotation/check_incorrect.txt: Lines where normalization changed text: original<TAB>normalized.
annotation/length_hist.jpg: Histogram of character lengths (generated by plot_length_histogram.py).
annotation/language_pie.jpg: Language mix pie chart (generated by plot_language_pie.py).

Citation

If you use this dataset, please cite:

@misc{soyvitou_kptma2026,
title        = {Khmer OCR Dataset Collection (KEPTA)},
author       = {Soy, Vitou and Eong, Koungmeng},
year         = {2026},
note         = {Benchmark dataset for Khmer OCR; arXiv preprint coming soon}
}

(Optional sections you can add later: “How to load in 🤗 Datasets”, evaluation protocol (CER/WER), license/ethics/data sources.)

Downloads last month: 3,190

Size of downloaded dataset files:

14.3 MB

Size of the auto-converted Parquet files:

14.3 MB

Number of rows:

3,383