johnlockejrr/abyssinica-synth

Abyssinica Synth

Synthetic line-level OCR/HTR training data for Amharic, Tigrinya, and Geʿez, rendered with manuscript-style page layouts and cropped to individual text lines.

Hugging Face: johnlockejrr/abyssinica-synth

Dataset summary

Task	Optical character / handwriting recognition (line images → text)
Split	train only (176,349 examples)
Features	image (line crop), text (ground-truth string)
Scripts	Ethiopic (Amharic, Tigrinya, Geʿez)
Generator	SvelteSynth Manuscript layout generator v0.0.1
On-disk size	~6.2 GB (images dominate)

Pages were generated as PNG + PAGE-XML pairs, then line regions were cropped from the page images using the PAGE-XML boxes. Each row is one line image and its Unicode transcription.

Generator

Fonts

Templates cycle through seven self-hosted Ethiopic fonts (not a single typeface):

File	Notes
AbyssinicaSIL-Regular.ttf	SIL Abyssinica
Brana-Regular-v1.0.0.ttf	Brana
EthiopicAbay-Regular.ttf	Abay
EthiopicLeTewahedo-Regular.ttf	Le Tewahedo
EthiopicLessan-Regular.ttf	Lessan
EthiopicSadiss-Regular.ttf	Sadiss
NotoSerifEthiopic-Regular.ttf	Google Noto Serif Ethiopic

Fonts were loaded via @font-face in the renderer so batch output is reproducible per template. Different batch runs / templates use different families from this pool.

Languages and text sources

Language	HF tag	Source material
Amharic	am	Genesis (Bible, book 1); plus history text from Amharic-English-Machine-Translation-Corpus — Amharic_English_History/amharic.txt
Tigrinya	ti	Genesis (Bible, book 1)
Geʿez	gez	Genesis (Bible, book 1); plus Ethiopic poetry from Malkea-Gubae (malkǝʾ and related hymnographic texts from the project’s published work records; see malkeagubae.com)

Corpus text was preprocessed (including optional Unicode normalization), then paginated through SvelteSynth batch generation. Layouts include single-column and multi-column flow templates, varied parchment-style backgrounds, and 300 DPI rasterization. Visual plausibility only — column contents are not semantically paired across regions.

How it was built

1. Corpus — UTF-8 .txt files per language/domain (see table above).
2. Layout — JSON templates in SvelteSynth (margins, 2-flow / 3-flow columns, justified LTR body, backgrounds from static/backgrounds/).
3. Rasterization — Headless Chromium; one canonical /render path for preview and CLI.
4. Ground truth — PAGE-XML (PAGE namespace 2019-07-15): TextRegion / TextLine with Coords, Baseline, and TextEquiv/Unicode.
5. Line export — Each visible line cropped to an image; paired with the PAGE-XML Unicode string → Hugging Face image + text columns.

Dataset structure

from datasets import load_dataset

ds = load_dataset("johnlockejrr/abyssinica-synth", split="train")

example = ds[0]
example["image"]  # PIL Image — one text line
example["text"]   # str — ground-truth transcription

• 176,349 training examples
• No dev/test split in this release — hold out your own for evaluation.

Intended use

• Training or fine-tuning line-level HTR/OCR for Ethiopic-script languages
• Augmenting real manuscript or print scans with layout/font diversity
• Not intended as parallel MT data (languages and sources are not aligned sentence-wise across the dataset)

Limitations

• Synthetic only — static background textures; no ink bleed, foxing, or binding curvature.
• Seven fonts — still far fewer than historical manuscript hands; expect domain shift on unseen typefaces.
• No per-row metadata in the published schema — language, font, template, and source are not separate columns; infer from your training recipe or add them in a future revision.
• Mixed domains — Amharic and Geʿez mixes biblical and non-biblical text; filter if you need a single genre.

Licensing

Component	License
This dataset (synthetic line crops + transcriptions)	Apache 2.0
Amharic-English-Machine-Translation-Corpus	MIT — contact author for commercial use
Malkea-Gubae	CC BY-SA 4.0
Noto Serif Ethiopic	SIL Open Font License
Abyssinica SIL	SIL Open Font License
Other Ethiopic font files	Check each font’s bundled license before redistribution

Respect upstream corpus and font licenses when redistributing or merging with other data.

Citation

@dataset{abyssinica_synth_2026,
  author       = {johnlockejrr},
  title        = {Abyssinica Synth: Synthetic Ethiopic Line Images for HTR},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/johnlockejrr/abyssinica-synth}}
}

Source corpora:

• MarsPanther, Amharic-English Machine Translation Corpus, GitHub.
• Augustine Dickinson, Malkea Gubae (መልክአ ጉባኤ), malkeagubae.com / GitHub.

Generator

SvelteSynth Manuscript layout generator v0.0.1 — form-based layouts, DPI-correct PNG + PAGE-XML, batch corpus pagination, then line cropping for upload.