ssdataanalysis/hebocrbench-v1

HebOCRBench v1 — Modern Hebrew participant pack

This is the public, fixed evaluation-input pack for the five-track HebOCRBench 1.0 Modern Hebrew printed-document headline. It contains 34,267 images and opaque input records. It intentionally contains no transcription gold, original source identifiers, source URLs, private filenames, organizer ID map, or HMAC key.

The unchanged certified references are withheld from this repository. An organizer remaps submitted opaque identifiers and scores predictions against those references. Withholding the gold is an evaluation protocol: the source material is public-fixed, so this release does not claim that models have never seen it or that it is contamination-free.

Exact release identity

Binding	Value
Benchmark version	1.0.0
Suite fingerprint	c68250ec4320485e243171b7d3f86c9b3b526f8ada317eda592cd7289f4df5ea
Participant-pack fingerprint	957e3b4f05707155db407d6707d742c7fe459365f575febb963e9d5732c3913c
Suite-lock SHA-256	04709dd2a76882fdb02704134dba3b3dfcc59b90dc967147f1a3b31d6d70c9ef
Locked files	34,275, plus PACK-LOCK.json
Image records	34,267
Sharding	two lowercase hex characters; maximum 139 images per directory

The byte-locked pack is under participant-v1.0.0/. Its public verification material is:

• PACK-LOCK.json: SHA-256 and size for every locked file, plus the pack fingerprint;
• PACK-MANIFEST.json: release, track, count, sharding, and no-leakage declarations;
• suite-lock.json: the exact five-track suite binding.

The gold_sha256 values in the two public manifests are one-way integrity bindings only. No reference text or reference structure is included.

Tracks

Track	Inputs	Mode
modern-bidi-v1	192	Unicode/BiDi conformance with supplied geometry
modern-line-recognition-v1	28,414	line recognition with supplied geometry and opaque line IDs
modern-page-ocr-v1	700	blind end-to-end page OCR
modern-tables-v1	61	blind table detection, structure, spans, positions, and cell text
modern-robustness-v1	4,900	blind OCR on locked degradation pairs

All Hebrew text in predictions must be in logical Unicode order. Do not mirror images or reverse Hebrew strings. Latin runs, numbers, URLs, email addresses, identifiers, and punctuation keep their ordinary internal order.

Download and verify

hf download ssdataanalysis/hebocrbench-v1 \
  --repo-type dataset \
  --local-dir hebocrbench-v1

cd hebocrbench-v1/participant-v1.0.0
jq -r '.files[] | "\(.sha256)  \(.path)"' PACK-LOCK.json \
  | shasum -a 256 --quiet -c -

The checksum command exits with status 0 only when every locked file has the declared bytes. PACK-LOCK.json itself declares the participant-pack fingerprint shown above.

Run a model and prepare predictions

For each track, read:

participant-v1.0.0/tracks/<track-id>/inputs.jsonl

Resolve each image.path relative to participant-v1.0.0/. Write one JSON object per input line to a separate prediction JSONL for that track. Copy the opaque page_id exactly and include schema_version: "1.0" and track_id.

At minimum, emit the fields required by the selected task:

• BiDi and line recognition: preserve the supplied opaque region/line IDs and add recognized text in their lines;
• blind page OCR and robustness: emit discovered regions, line text, and reading_order;
• tables: emit detected tables and their cells, spans, positions, geometry, and logical-order cell text.

Example envelope (the exact nested fields depend on the track):

{"schema_version":"1.0","track_id":"modern-page-ocr-v1","page_id":"hbo-v1-p-…","regions":[],"reading_order":{"edges":[]},"tables":[]}

A submission must contain exactly one prediction for every page in the track; unknown, duplicate, or missing opaque page IDs are rejected. The prediction schema, track contracts, organizer remapping tool, official scorer, and reproducibility documentation live in the HebOCRBench code repository.

What this pack does not cover

This release is only the five-track Modern Hebrew print headline. It does not blend scientifically different families into one score. The full suite also records separate real-data extensions for writer-disjoint modern handwriting, a narrow six-page Pinkas handwriting subset, and 34 pages of mixed square/Rashi HaZefira historical press. Synthetic niqqud and single-font Rashi probes are non-rankable diagnostics.

The following remain explicit missing real-data coverage in v1:

• forms with field-level ground truth;
• Biblical Hebrew with real cantillation ground truth aligned to scans;
• pure-Rashi historical print with region- or line-level Rashi labels.

No score from this pack should be presented as a handwriting, Biblical, cantillation, historical-print, or pure-Rashi result.

Provenance and terms

HebOCRBench software is MIT-licensed. That code license does not relicense the images in this dataset. Each source document retains its own terms, rights, and attribution requirements. The benchmark build records revision-locked acquisition evidence, hashes, page counts, provenance, and source-specific terms; upstream identities and URLs are deliberately omitted from this blind participant pack. Consult the code repository's source and license audit before redistributing or using source images beyond benchmark evaluation.

Citation

Cite the HebOCRBench software and report the exact suite, track, dataset, and participant-pack fingerprints used for every result. The canonical citation metadata is maintained in CITATION.cff in the code repository.

---

עברית בקצרה: זהו מארז הקלט הציבורי והנעול של חמשת מסלולי הדפוס העברי המודרני. ה־gold נשמר בנפרד, המזהים אטומים, וכל תוצאה חייבת לציין את טביעות האצבע המדויקות של ה־suite וה־pack. אין להציג תוצאה ממנו כתוצאת כתב־יד, מקרא עם טעמים או Rashi היסטורי טהור.