heb-words100k-uni10aug-diffpen

DiffusionPen finetune data at word level: 100,000 unique Hebrew words, each written once in one of 10 handwriting styles (participant_01 to participant_10; label 0-9). Same words, same style per word and same train/test split as cyttic/heb-words100k-uni10-diffpen; only the drawing differs.

• Cursive spacing: each letter is placed against the ink of the letters already drawn, row by row. In the letter-body zone the whole body of the neighbour blocks (no letter slips into a bowl); above and below it the outline is followed, so ascenders and descenders may reach over / under a neighbour.
• Tail rule: for נ ת כ צ, when the glyph has a bottom tail running left of its body (detected per glyph image), with probability 0.7 the next letter is placed against the body only and sits on the tail. Applied 44,366 times (ת 23,878, נ 14,117, כ 4,018, צ 2,353).
• Augmentation: word size x U(0.88, 1.12), slant (shear) U(-0.22, 0.22), word rotation U(-2.5, 2.5) deg, stroke thickening k in [0, 1, 2] px at 4x (p = [0.45, 0.35, 0.2]); per letter a random one of the writer's 3 handwritten repetitions, scale x U(0.88, 1.12), rotation U(-6, 6) deg, baseline shift U(-0.06, 0.06) x size, gap U(0.5, 6.0) px.
• Layout: 256x64 RGB, white page, black ink, EM 40 px on baseline y = 38, ink centred; 0 words shrunk uniformly to fit 248 px; 389 words have some ink cut at the top / bottom page edge.
• Split: 97,979 train / 2,021 test. Columns and files as in the plain set.
• Built by diffpen-hebrew2/build_word_ds_uni10aug.py.