cyttic/diffusionpen-hebrew-handwriting-cer0

DiffusionPen Hebrew Handwriting — CER=0 clean subset

The highest-fidelity slice of DiffusionPen Hebrew Handwriting: only the 33,082 line images that an independent Hebrew TrOCR read back with an exact match (character error rate = 0.0). Built to test whether a smaller, label-clean set trains a better recognizer than the full (noisier) 150k set.

• 33,082 line images, 491 writer styles
• Splits (writer-independent, style-disjoint): train 26,607 / validation 3,187 / test 3,288
• Every row had line CER == 0.0 under the independent TrOCR audit

Fields

field	type	description
image	image	RGB PNG, 64 px tall — one line of Hebrew handwriting
text	string	ground-truth transcription
style	int32	writer-style id [0, 490]
cer	float32	independent TrOCR line CER — 0.0 for every row here

Caveat

CER=0 means the audit TrOCR transcribed the line perfectly — the cleanest available signal, but not a guarantee of zero label noise: TrOCR's language model can read through a subtly malformed glyph and still report 0.0. This is the best-fidelity subset, not a certified-perfect one; cross-checking with a second recognizer is recommended for the strictest use.

See the parent dataset card for the full generation, audit, and refinement methodology.