cyttic/pinkas-hebrew-lines

Pinkas — Hebrew cursive text lines (line crops)

Line-level crops of The Pinkas Dataset: 30 pages of early-modern European Jewish community records (c. 1500–1800), written by non-professional scribes in Hebrew and Yiddish, with line-level ground truth in PAGE XML.

Original dataset — all credit to its authors: Irina Rabaev, Berat Kurar Barakat, Jihad El-Sana. The Pinkas Dataset. Zenodo, 2019. https://doi.org/10.5281/zenodo.3569694 (record 3569694). License: CC-BY-4.0. This is a derived, re-formatted copy under the same license; please cite the original.

What was done to the original

• Each TextLine of the PAGE XML is cut out along its polygon; pixels outside the polygon are filled with the page's own background level, so neighbouring lines do not leak into the crop, then the crop is tightened to the kept ink (parsing and cropping code from itayinbar/Mishkefet-v1, hebocr/data/hebrew_ink.py).
• Text is normalized (geresh/gershayim to ASCII quotes, whitespace collapsed); lines with fewer than 2 characters or under 50% Hebrew letters are dropped.
• Each line is stretched horizontally toward 22.9 pixels per character at 64 px height (the median density of modern Hebrew handwriting lines in cyttic/trocr-hebrew-matan); Pinkas is natively ~2.3x denser. Long lines are capped at 2,560 px, so the median reaches ~16 px/char.
• 5 lines too wide for a 2,304 px budget at 64 px height were dropped.

Fields

field	description
image	grayscale PNG line crop
text	transcription (normalized)
writer	pinkas:<page name> (page-level id; scribes are not annotated)
orig_split	train / test per the original train_set.txt / test_set.txt

938 lines from 27 pages (266 of them from the original test pages).