Khattat

Synthetic multilingual handwritten OCR dataset by saeidseyfi.

Handwritten samples in Persian (fa), Arabic (ar) and English (en) rendered with a dynamic pen (variable stroke width and pressure along the writing path), with three configs:

Configs

1. default — line crops (train 21,726 / validation 2,062 / test 924)

Handwritten line images with ground-truth transcriptions.

Column	Type	Description
image	image	RGB handwritten line crop
text	string	ground-truth transcription (original Unicode, incl. ZWNJ)
id	string	unique sample id
language	string	fa / ar / en
style	string	handwriting style bucket (light / medium / heavy)
augmented	bool	true for augmented variants (scanner/camera distortions)

Train: 7,242 clean + 14,484 augmented. Language mix: fa 10,533 · ar 6,627 · en 4,566. Splits are writer-disjoint (no handwriting style appears in more than one split).

2. pages — full pages with formulas & shapes (train 199 / validation 22 / test 19)

Full handwritten pages containing language text, mathematical formulas (arithmetic, linear equations, fractions, square roots, percentages, trigonometry, integrals, series — all mathematically true) and hand-drawn shapes (circle, ellipse, square, rectangle, triangle, pentagon, hexagon, star, arrow, line) with pen-pressure strokes, plus caption annotations (e.g. r = 5 cm).

Column	Type	Description
image	image	full page image
page_id	string	page identifier
language	string	primary page language
page_type	string	text / mixed / math
full_text	string	all text lines + formula labels in reading order
n_lines	int	number of text lines
n_formulas	int	number of formulas
n_shapes	int	number of drawn shapes
shapes_label	string	shape names, e.g. circle; triangle
regions	string	JSON with bounding boxes & labels for every element: header, page number, text lines, formulas, shape regions

3. formulas — formula line crops (train 371 / validation 41 / test 29)

Individual formula crops labeled with their exact transcription (e.g. 3/4 + 1/6 = 11/12, sin(30°) = 0.5, √64 = 8).

Column	Type	Description
image	image	formula crop
text	string	exact formula transcription
template	string	formula template family
page_id	string	source page

Splits for pages / formulas are page-disjoint.

Usage

from datasets import load_dataset

lines = load_dataset("saeidseyfi/khattat")                    # default: lines
pages = load_dataset("saeidseyfi/khattat", "pages")           # full pages
forms = load_dataset("saeidseyfi/khattat", "formulas")        # formula crops

example = pages["train"][0]
print(example["full_text"])
example["image"].show()

# clean-only line training
clean = lines["train"].filter(lambda x: not x["augmented"])