The dataset viewer is not available for this split.
Parquet error: Scan size limit exceeded: attempted to read 309745881 bytes, limit is 300000000 bytes
Make sure that
1. the Parquet files contain a page index to enable random access without loading entire row groups2. otherwise use smaller row-group sizes when serializing the Parquet files
Error code: TooBigContentError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
heb-words100k-uni10-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 of the letter-scan forms = the styles of the exp39 human10uni sets; label 0-9). Words were shuffled
and dealt round-robin to the styles, so each style has ~10,000 words.
- Words: the 100,000 most frequent words (2-10 letters, letters only, final letters only at the end) of the BYP and LLMGen corpora (counts summed; least frequent word kept occurs 4 times).
- Augmentation (the exp36 / exp39 unigram jitter): word size x U(0.95, 1.05); per character a random one of the writer's 3 handwritten repetitions, scale x U(0.94, 1.06), rotation U(-3, 3) deg, baseline shift U(-0.03, 0.03) x size, advance x U(0.95, 1.08).
- Layout: 256x64 RGB, white page, black ink, fixed letter size (EM 40 px, typical letter ~22 px) on a fixed baseline (y = 38), ink centred horizontally; no crop-and-stretch. 0 words were shrunk uniformly to fit 248 px.
- Split: by word, 97,979 train / 2,021 test.
- Columns:
image,text(logical order),writer,label,relpath(<writer>/word_<frequency rank>.png),kind,split. - Files:
data/train.parquet,data/test.parquet,writers_dict_*.json,character_classes.json,letter2index.json,index2letter.json,words.tsv(rank, word, corpus count, writer, split),stats.json,preview/. - Built by
diffpen-hebrew2/build_word_ds_uni10.py.
- Downloads last month
- 78