cyttic/heb-synth-punct-digits-hhf

Hebrew synthetic handwriting lines: digits, gershayim, punctuation

20,000 synthetic Hebrew text lines built to teach a line recognizer the characters that real Hebrew handwriting datasets barely contain: digits, gershayim/geresh, quotes and punctuation.

How it was made

• Handwriting: fonts built from cropped real handwriting — 4 participants filled scanned templates in which every character (27 letters, 10 digits, ! ( ) , - . / : ; ? ׳ ״) was written 3 times; one font per repetition (12 fonts).
• Per line one participant; per character one of that writer's 3 handwritten versions at random, plus jitter (scale ±6%, baseline ±3%, rotation ±3°, advance ±, word gaps ±).
• Layout: Unicode bidi (python-bidi): Hebrew right-to-left, numbers left-to-right, brackets mirrored; lines drawn character by character in visual order.
• Text: modern Hebrew sentences from an LLM-generated corpus (DictaLM 2.0 / GPT-4.1-nano), selecting every line with an in-word gershayim or geresh, then lines with digits, then other punctuation.
• Gershayim: the templates contain handwritten ״ / ׳ but no ASCII quotes, so the handwritten gershayim is drawn for both roles (abbreviations like קמ"ש and quotation marks).
• 64 px line height; width matched to ~22.9 px per character (±random).

Labels

text uses ASCII " and ' for gershayim/geresh (the canonicalization used by the ivrit.ai benchmark scoring). style = participant id.

content	count
lines	20,000 (≈5,000 per writer)
lines with in-word gershayim	4,325
lines with digits	13,080