A small, human-verified benchmark for OCR / handwritten text recognition (HTR) on modern Hebrew handwriting: 225 gold lines across 10 pages, one page per writer, drawn from the transcriptor.ivrit.ai volunteer transcription corpus.

This is a test set. There is no train split, by design — it exists to be held out. It is deliberately small and clean rather than large and noisy: every line was transcribed by at least two volunteers independently and then verified.

Results for current frontier vision models are on the leaderboard Space.

Configs

from datasets import load_dataset

lines = load_dataset("ivrit-ai/hebrew-handwriting-ocr-benchmark", "lines", split="test")   # 225 rows
pages = load_dataset("ivrit-ai/hebrew-handwriting-ocr-benchmark", "pages", split="test")   # 10 rows

config	one row is	`image`
`lines` (default)	a gold text line	the cropped line, full resolution, upright
`pages`	a page	the whole page, full resolution, upright, with all its gold lines nested under `lines`

The two mirror the two ways the benchmark is run: feed the model one line at a time, or feed it a page and ask for everything on it.

Coordinate contract

Every bbox and polygon is in the pixel grid of the image in the same row. Cropping is just:

page = pages[0]
x, y, w, h = (page["lines"][0]["bbox"][k] for k in ("x", "y", "w", "h"))
line_image = page["image"].crop((x, y, x + w, y + h))

No rotation handling, no rescaling. image_rotation is recorded for provenance but has already been applied — the published images are upright. The untouched upstream coordinates (in the source corpus's downscaled-jpg grid, which is not the grid of the stored image on rotated pages) are kept as bbox_jpg / polygon_jpg for traceability only.

How the gold text was made

Volunteers transcribed detector-proposed lines on transcriptor.ivrit.ai. Only lines with ≥2 independent transcriptions are eligible.
Text was canonicalized — quotes, dashes, the Hebrew maqaf / geresh / gershayim, ellipsis, Unicode spaces, zero-width and bidi marks — leaving Hebrew letters and niqqud untouched.
Lines where the volunteers agreed exactly were auto-accepted; the rest went through human adjudication against the line crop.
Pages were chosen status-clean, one per writer fingerprint.

Fields

`lines`

image, text (the gold), line_id, external_id, page_id, submission_id, line_index, bbox, polygon, bbox_jpg, polygon_jpg, n_votes, decision, chosen_read_index, reads, transcriptions (the raw volunteer records: user_id, kind, text, timestamps), detection_confidence, image_rotation, page_width, page_height, source, upstream_license (the source corpus's per-batch licence, not this dataset's), plus the writer block below.

`pages`

image, page_id, page_text (gold lines joined by newline, in reading order), lines (list of line_id / line_index / text / bbox / polygon / n_votes / decision / detection_confidence), n_lines, image_width, image_height, jpg_width, jpg_height, image_rotation, submission_id, page_external_id, document_name, source, upstream_license, plus the writer block.

Writer block (both configs)

decade_written, writer_age_range, writer_gender, native_language, legibility_score (1–10, self-reported), consent_given, additional_notes, submitter_fingerprint. Useful for slicing results by hand or legibility.

Suggested metrics

The reference harness reports, over normalized text:

Word coverage (headline) — the multiset fraction of a gold line's tokens present in the model's output. Order- and segmentation-insensitive, so it works unchanged for the pages config, and a verbose hallucination cannot inflate it the way it deflates CER.
CER — median (the typical line) and micro (total edits / total reference characters). Mean CER is not reported: a few runaway outputs dominate it.
Blank outputs are excluded from both and counted separately.

Reference implementation: run_ocr_eval.py in the ocr-eval harness.

Limitations

Small. 225 lines / 10 writers — enough to rank models coarsely, not enough for tight confidence intervals or fine-grained error analysis.
Detector-proposed boxes. Lines are initially identified by an automatic detector and then manually corrected, so some bounding boxes are loose and a few overlap neighboring lines. The gold text corresponds to the text of that line.
Loose crops on one page. 17 lines sit on a 180°-rotated page whose boxes are unusually tall; they render upright and their gold text was re-verified against correctly-oriented crops for this release.
Domain skew. Mostly personal notes, school notebooks and forms, written in ballpoint on lined or blank paper, photographed with phones.

Personal and sensitive information

Rows carry self-reported writer demographics (age range, gender, native language, decade written), a submitter_fingerprint, and the volunteer user_ids who transcribed each line. All of it was contributed under explicit consent and is already public in the upstream corpus. The manuscript images themselves may contain incidental personal content written by their authors.

License and citation

The dataset is released under the ivrit.ai License, which enables broad research and commercial use.

Full license: https://www.ivrit.ai/en/the-license/
FAQs: https://www.ivrit.ai/en/license-faqs/

@misc{hebrew_handwriting_ocr_benchmark,
  title  = {Hebrew Handwriting OCR Benchmark},
  author = {ivrit.ai},
  year   = {2026},
  url    = {https://huggingface.co/datasets/ivrit-ai/hebrew-handwriting-ocr-benchmark}
}

Downloads last month: 148

Total file size:

23.3 MB