heb-words27-diffpen

Word-level Hebrew handwriting for DiffusionPen finetuning -- the step up from cyttic/heb-connected-bigrams17-diffpen, which covered 2-character units. Here the unit is a whole word.

Content

• 20,000 most frequent words of LLMGen2 (sentences_llm2.txt, 968,204 LLM-generated Hebrew sentences, 167,879 unique words)
• 643 numbers -- LLMGen2 contains no digits at all, because the generation prompt forbade them, so numerals had to be added separately or the model could never write one
• 27 styles = 9 participants x 3 repetitions of the handwriting collection form, traced to TTF by letter-scan-extractor. Each participant-repetition is its own style class (style_classes=27).
• 557,361 images, 256x64 RGB -- pixel-compatible with the bigram dataset (identical normalize()), so a model trains on either without a preprocessing switch.

Columns

image, text, writer (e.g. p04r2), label (int style class), relpath, kind (word / number), split.

Word relpath encodes frequency rank (word_000123.png), so taking a prefix gives the N most common words rather than a random slice.

Split

By word, not by image. A word is train or test in every style. A per-image random split would place the same word on both sides in different hands, letting the content encoder memorise rather than generalise. ~4% test, hash-based and reproducible. Verified: zero word overlap between the splits.

Vocabulary

letter2index.json extends the Hebrew set with the 10 digits -- existing indices kept stable, digits appended -- so a warm start from a Hebrew checkpoint still lines up on the characters it already knows. 40 characters total.

Provenance

Text: GPT-4.1-nano via the OpenAI Batch API (openai_batch_sentences_refined.py). Styles: letter-scan-extractor -> build_font.py (potrace outlines, one shared scale factor, per-character baseline placement). Build: diffpen-hebrew2/build_word_ds27.py.