heb-bigrams-new10-diffpen

DiffusionPen finetune data for 10 new handwriting styles (nothing else: no older fonts).

• Styles: pt01 to pt10 = participants 01-10 of the letter-scan forms (letter-scan-extractor); style slots 0-9 (label column).
• Units: every Hebrew bigram (22 base letters x 27 letters, final letters only as the second letter = 594) plus every single letter in each of the participant's 3 variants (27 letters x 3 = 81 images) = 675 images per style, letters only (no digits, no punctuation). One image per (style, unit): 6,750 images (6,480 train / 270 test, random split).
• Rendering: each participant wrote every letter three times on the form. Bigrams: each letter of each image is drawn at random from that participant's three variants (variants_used.tsv records the choice), right to left on a shared baseline. Single letters: each letter once per variant (pure repetition 1, 2, 3, not mixed). 256x64 RGB, white page, black ink, ink centred, a typical letter about 22 px tall (same layout as cyttic/heb-bigrams17-diffpen).
• Columns: image, text (logical order, first letter = rightmost), writer (pt01..pt10), label (0..9), relpath, split.
• Files: data/train.parquet, data/test.parquet, writers_dict_*.json, character_classes.json, letter2index.json, index2letter.json, variants_used.tsv, stats.json, preview/.
• Built by diffpen-hebrew2/build_bigram_ds_new10.py. Intended use: finetune the ORIGINAL DiffusionPen (IAM-pretrained, no Hebrew stage) on these 10 styles.