mr3vial/paleo-hebrew-seals-unambiguous

PaleoHebrew-Seals Real Benchmark (Unambiguous Subset)

This repository hosts the real benchmark part of PaleoHebrew-Seals, a dataset suite for multimodal recognition of Paleo-Hebrew seal inscriptions from photographs.

Why this dataset is needed

Paleo-Hebrew seal inscriptions are difficult for standard OCR systems: the signs are sparse, shallow, frequently worn, and embedded in irregular seal impressions captured under uncontrolled lighting and viewpoint changes. Off-the-shelf Hebrew OCR systems are optimized for modern square Hebrew and perform very poorly on this domain. This benchmark provides structured supervision on real seal photographs for localization, classification, transcription, and glossing.

Source and subset definition

The benchmark is derived from the public Hebrew Stamp Seals (HSS) interface hosted through the University of Bern nodegoat system.

This release is intentionally restricted to an unambiguous subset. A real sample is included only if both conditions are satisfied:

1. the corresponding source entry contains a deciphered Hebrew inscription; and
2. the image contains at least one visible sign that can be localized with a bounding box.

This rule produced 350 candidate images, of which 307 remained in the final benchmark after box-level inspection.

What each sample contains

Each benchmark sample contains:

• a seal photograph
• a sequence of character labels
• aligned character-level bounding boxes
• Hebrew text variants
• transliteration
• English gloss
• selected catalog metadata
• split assignment

When multiple photographs correspond to the same seal entry, the release preserves a row_id field so that grouping remains explicit.

Splits

The current split contains:

• 157 training images
• 150 validation images

Whenever multiple photographs belong to the same underlying seal entry, split assignment is defined at the seal-entry level.

Tasks supported

This benchmark supports four core tasks:

1. character localization
2. character classification
3. Hebrew transcription
4. English glossing

Representative fields

Representative fields in the release include:

• image
• row_id
• chars
• bboxes
• Hebrew text variants
• transliteration
• English gloss / translation
• selected catalog metadata
• split

Baseline anchors

Representative anchors reported for the real validation split:

Localization and crop classification

Method	Precision	Recall	mAP50	mAP50-95
YOLO detector	0.942	0.894	0.947	0.608

Method	Accuracy
ConvNeXt classifier	0.726

Transcription and glossing

Task	Method	CER	WER	chrF	EM
OCR	Tesseract (heb)	3.400	7.029	1.817	0.000
OCR	Kraken MiDRASH_Gen_01	2.832	5.294	1.811	0.000
Transcription	Detector + classifier top-1	0.350	0.905	31.001	0.073
Transcription	mT5-large post-OCR	0.353	0.642	45.150	0.247
Transcription	Qwen3-VL LoRA (mask)	0.485	0.697	42.122	0.340
Gloss	mT5-large translate	0.344	0.424	60.690	0.313

These numbers are included as release anchors rather than saturation points. They illustrate the domain gap between standard Hebrew OCR and Paleo-Hebrew seal imagery.

Split policy and leakage control

Leakage control matters in this setting. In the release protocol, any real images used for Stage B style adaptation in the synthetic pipeline are treated as training-only resources and are kept disjoint from benchmark evaluation artifacts at the seal-entry level.

Intended use

This dataset is intended to support research in:

• multimodal OCR
• historical document analysis
• ancient-script recognition
• digital humanities
• cultural-heritage computing
• synthetic-to-real transfer

It is the primary resource for evaluation on real seal photographs.

Limitations

This benchmark is intentionally restricted to an unambiguous subset. This improves label reliability, but it does not cover the full ambiguity spectrum of historical epigraphy. Automatic predictions on seal imagery should not be treated as authoritative scholarly readings. The dataset is meant to support expert analysis and reproducible research, not to replace epigraphic judgment.

Companion resources

• Synthetic corpus: mr3vial/paleo-hebrew-seals-synthetic
• Demo Space: https://mr3vial-paleo-hebrew-project.hf.space/
• Demo video: https://drive.google.com/file/d/1susDDbaZyFny1Ga9bZXyEVibD4R8YyrW/view

Access

This repository is intended to be publicly accessible without login and without access requests.

License

The dataset contents in this repository are released under CC BY 4.0.

Companion code, evaluation scripts, and model checkpoints may be documented and licensed separately in their respective repositories.

Citation

If you use this resource, please cite the dataset paper as follows while the submission is under review:

@misc{gorbulev2026paleohebrewseals,
  title={PaleoHebrew-Seals: A Real-and-Synthetic Dataset Suite for Multimodal Recognition of Paleo-Hebrew Seal Inscriptions},
  author={Gorbulev, Alex and Humonen, Innokentiy and Golyadkin, Maksim and Makarov, Ilya},
  year={2026},
  note={Under review}
}

Contact

For questions about the benchmark, please contact the repository maintainers.