Datasets:

jakeboggs
/

HieroglyphBench

license: other
license_name: glyphdataset-non-commercial
pretty_name: HieroglyphBench
task_categories:
  - image-to-text
tags:
  - ocr
  - hieroglyphs
  - egyptology
  - gardiner
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: test
        path: test/*

HieroglyphBench

An OCR benchmark for ancient Egyptian hieroglyphs: 30 photographed columns of hieroglyphs from the Pyramid of Unas (11-26 signs each), each paired with an expert transcription of the signs as Gardiner sign-list codes. The task is to read the signs in the image, in order.

Writeup and leaderboard: https://boggs.tech/posts/reading-hieroglyphs-with-vlms/

Code and evaluation harness: https://github.com/JakeBoggs/HieroglyphBench

Examples

Photo	Ground-truth signs
	_{D21 G17 O4 D21 G43 N5 Q3 N35 G17 M17 D4 E23}
	_{D46 V4 N14 S29 N35 V13 G43 G17 D21 N35 V31 G43 Q3 N35}
	_{G17 E34 N35 M17 S29 M18 M17 G17 E34 N35 M17 S29 M18 M17 G17 V31 E34 N35 M17 S29}

Columns

image - the cropped photograph of the inscription column
id - item id (unas-p{plate}-c{col}-{chunk})
plate - source plate and column
num_signs - number of ground-truth signs
gardiner - ordered list of Gardiner codes (the ground truth)
gardiner_str - the same codes, space-joined
unicode - the signs rendered as Unicode hieroglyphs (U+13000 block)

Source and license

Photographs and per-sign annotations come from Morris Franken's GlyphDataset, derived from Alexandre Piankoff's The Pyramid of Unas (1955) plates. Non-commercial use; keep the Franken / Piankoff attribution.