MendyA/snapgrade-synth-v1

SnapGrade synthetic Hebrew exam handwriting (v1)

These are synthetic photos of handwritten Hebrew exam answers, with exact transcriptions. They were made to train and test handwriting readers for SnapGrade, an app that helps teachers grade handwritten exams.

Every image was drawn by our generator (ocr-lab/synth/generate.py). Because we drew it, we know exactly what it says. The generator works in four steps:

1. Text: Hebrew answer sentences, with deliberate spelling mistakes added.
2. Ink: the words are rendered from handwriting-style fonts, real handwritten letters, or whole DiffusionPen lines. Each "writer" gets their own slant, size, spacing and wobble.
3. Layout: answers go on exam pages, sometimes with columns, margin continuations, words inserted above the line, and crossed-out words.
4. Photo: the page goes on plain, lined or grid paper in pencil, blue or black ink. Then we add phone-photo effects: uneven light, shadow, blur, tilt and JPEG compression.

Contains no real pupil data.

Contents

Each <folder>.tar unpacks to <folder>/images/*.jpg plus <folder>/manifest.jsonl. Checksums are in SHA256SUMS.

Folders	Samples	Purpose
lines_0 – lines_3	10,000 single lines	training
regions_10 – regions_12	6,000 answer regions (one answer, several lines)	training
pages_20 – pages_25	3,000 full exam pages	training
regions_val, pages_val	300 regions, 100 pages	validation
tricky_test	100 pages with every hard layout turned up; held-out letters and print font	test
regions_text_seed, regions_text_heldout, regions_text_random	3 × 150 regions with the same writers and layouts, but different text: training sentences / unseen sentences / random-letter words	test: does a reader read, or guess from memory?

Each manifest line has:

• id, image (relative path) and text, the exact transcription.
• level: line, region or page.
• split: train, val or test.
• writer: the handwriting source.
• misspellings: the injected spelling mistakes, which a reader must copy, not fix.
• Page records only: regions, the box and text of each answer.

Markup inside text:

• \n means a new line, and <col> a new column.
• <ins>…</ins> marks a word inserted above the line.
• <del>…</del> marks crossed-out text.

Use it

Easiest, from the SnapGrade repo: python ocr-lab/setup/assets.py fetch synth-v1. That downloads, checks and unpacks every folder.

Or download the tar files directly:

from huggingface_hub import hf_hub_download
import tarfile
tar = hf_hub_download("MendyA/snapgrade-synth-v1", "regions_val.tar", repo_type="dataset")
tarfile.open(tar).extractall("data", filter="data")

Known limits

• Synthetic only. Real children's handwriting is messier than anything here. Use this data for pre-training and for checking layouts, not as proof of real-world accuracy.
• Small sentence pool. Train, validation and the first two test sets draw on the same ~125 answer sentences. A reader trained here can memorize them, so scores on those sets are optimistic. regions_text_heldout and regions_text_random exist to measure that gap.
• The cursive font is shared between training and tests; we have only one. Held out from training are the real handwritten letters (test split), the Amatic SC print font, and DiffusionPen's test writers.

License and credits

Released under CC-BY-4.0. It's built from:

• DiffusionPen Hebrew handwriting, CC-BY-4.0: whole synthetic lines.
• Hebrew handwritten letters (sivan22), CC-BY-3.0, derived from HHD_v0 (CC-BY-4.0): single letters.
• Google Fonts: Gveret Levin, Playpen Sans Hebrew, Amatic SC, Karantina, Varela Round, Noto Sans Hebrew, Rubik, David Libre, Frank Ruhl Libre and Heebo, all under the SIL Open Font License 1.1.
• Answer sentences written for this project.

If you use this data, please credit SnapGrade and the sources above.