Thareah/khmer-handwriting

Khmer Handwriting Lines

Images of handwritten Khmer text, one text line per image, each with its transcription. Use it to train and test handwriting recognition (HTR / OCR) models for Khmer.

Each line was marked with a bounding box and transcribed in a labeling app. The image is exactly that box, cut from the scanned page.

Splits

Split	Pages	Lines
train	118	2,311
validation	15	291
test	14	289
total	147	2,891

Splits are made by page, so lines from one page never appear in more than one split.

Columns

Column	Type	What it is
image	image	Grayscale image of one handwritten line, cut at its labeled box. About 111 px high; width varies with the line.
text	string	The Khmer transcription of the line.

Files

• train/, validation/, test/: the line images, plus a metadata.jsonl in each folder with the text for every image.
• charset.txt: all 117 characters that appear in the labels.
• stats.json: line and page counts per split, and how many annotations were dropped and why.

Label cleaning

Labels are normalized to Unicode NFC. Zero-width spaces are removed and runs of whitespace are collapsed to a single space. Boxes with empty text (7) or with multi-line or over-long text (19) were dropped.

Load it

from datasets import load_dataset

ds = load_dataset("Thareah/Tha_KhmerHandWriting")

row = ds["train"][0]
row["image"]  # PIL image
row["text"]   # Khmer transcription