heb-webfont-refoyl-byp

Synthetic Hebrew text-line images, one web handwriting font (Refoyl) x one corpus (BYP -- Ben-Yehuda Project literary Hebrew, punctuation kept, length-matched to LLMGen (sentences_byp_punct_250k_lenmatched_fit.txt)).

Part of a 10-font x 2-corpus grid (20 datasets, cyttic/heb-webfont-<font>-<byp|llmgen>). Within a corpus, all 10 font datasets hold the same lines in the same row order -- they differ only in the font. The few corpus lines too wide for some font (> 2304px at 64px; 3,633 BYP / 439 LLMGen, almost all from Anka CLM Bold) were replaced before rendering by another corpus line with the same word count that fits every font, so every dataset has the full 250,000 lines.

rows	250,000
columns	image (PNG, grayscale), text (verbatim label, punctuation + digits), style (font)
image	whole line normalized to 64px height, width <= 2304px, no augmentation
split	train only

Rendering: 48pt nominal, 6px pad, RTL shaping (raqm), seed 42. Fonts lacking a glyph for a corpus character were patched beforehand (patch_webfonts10.py): geresh/gershayim drawn with the font's own ' / "; Nachlaot . added as a dot of its own stroke width; Refoyl / borrowed from Anka CLM Bold. Labels are never altered.