maclenn77/ticha-htr-lines

Ticha manuscript lines (Colonial Zapotec and Spanish)

Line images of 17th to 19th century manuscripts from Oaxaca, Mexico, paired with their transcriptions, for training and evaluating handwritten text recognition (HTR). The pages and transcriptions come from Ticha: a digital text explorer for Colonial Zapotec (Haverford College).

The code that downloads the pages, pairs the lines and builds this dataset is at github.com/Maclenn77/ticha.

Configs and splits

Config	Split	Lines	Description
original	train	5,786	Original line images
original	validation	819	Original line images
original	test	1,066	Original line images
augmented	train	23,144	Augmented copies of the train lines (half degrade, half variety)
augmented	test_degraded	1,066	One degrade copy of each test line, no geometric changes

The splits are made by document: no document appears in two splits. Within each language, documents are assigned so that lines are close to 80/10/10 (7,671 original lines in total).

Augmented data is a separate config so it is never mixed in by accident. To train with it, concatenate original/train and augmented/train; validate on original/validation. Evaluate on original/test (real scans) and on augmented/test_degraded (the same lines made to look like low-quality scans) to see how much a model degrades on poor reproductions.

from datasets import load_dataset, concatenate_datasets

original = load_dataset("maclenn77/ticha-htr-lines", "original")
augmented = load_dataset("maclenn77/ticha-htr-lines", "augmented")
train = concatenate_datasets([original["train"], augmented["train"]])

Examples

Lines from the original config:

Image	Text	Language
	testamento suyo del mismo difunto su Marido Balthazar	Spanish
	Bixose copabijtoo Nijateni goca chij Sijganij xipe	Zapotec
	da Con Migel Luís tierra de lavichii es el Lindero Guetela	Spanish with Zapotec place names

The same line (SDE746dT_p001_l024) and two of its copies in the augmented config:

Version	Image
original
variety
degrade

Fields

original:

• id: <doc_id>_p<page>_l<line>
• image: line image (PNG), cropped along the line's polygon with the outside painted white
• text: transcription of the line, as in Ticha, including editorial marks such as q[ue], (?) and [roto]
• doc_id, page, line: Ticha document id, page number in the document, line number on the page
• language: Zapotec, Spanish or Spanish with some Zapotec
• source: auto (paired automatically) or reviewed (page checked line by line)
• year, town, doc_type, document_url: document metadata from Ticha

augmented: id, image, text, doc_id, page, line, language, plus source_id (the id of the original line), preset (degrade or variety) and params (JSON with every augmentation parameter).

How it was made

1. Page images were rebuilt from Ticha's zoom tiles; only documents with a transcription marked "Completed and Checked" or "Completed but not Checked" were kept (125 documents, 385 pages).
2. Pages were segmented into lines with kraken's default baseline model.
3. Each line was read by an HTR model (McCATMuS fine-tuned on Ticha lines) and the transcription was cut into the word spans that best match the readings, which also handles transcriptions whose line breaks do not follow the manuscript and words broken across lines.
4. Pairs were kept when their text fits the reading clearly better than the neighbouring lines' texts. Pages with too many doubtful pairs (48) were checked line by line against the images; lines that could not be confirmed have no text and are not in the dataset.
5. Augmentation: degrade imitates Ticha's untranscribed notarial scans (lower resolution, greyish paper with a cool cast, faded brown ink, washed-out or underexposed tone, uneven light, stains); variety adds brightness, contrast, blur, noise, JPEG compression, small rotation, slant, scale and elastic distortion.

Code: Maclenn77/ticha (ticha-scraper lines, scripts/build_dataset.py, scripts/export_hf_dataset.py).

Limitations

• Transcriptions follow Ticha's editorial conventions; some are partial or uncertain (marked with (?)).
• Alignment is automatic. Spot checks found the large majority of pairs correct, but a few lines may contain a word that belongs to the neighbouring line.
• 64 lines on reviewed pages were trimmed or corrected from the images by an AI assistant (Claude); the Zapotec ones in particular have not been checked by a specialist.
• Three characters (½, à, ò) appear in validation but not in train.
• The fine-tuned model used for alignment was trained on lines of all splits; do not use it to report test results on this dataset.

Licence

This dataset is a derivative of Ticha and is shared under the same licence, CC BY-NC-SA 4.0: non-commercial use only, with attribution, and any derivative must use the same licence.

Changes made to the Ticha material: page images rebuilt from the zoom tiles and cut into line images along each line's polygon; transcriptions split into lines and aligned with the images, with editorial line references (e.g. 63r-5) and pasted Word markup removed; 64 line texts on reviewed pages trimmed to what is visible on the line or corrected from the image; and, in the augmented config, synthetic copies of the line images.

Citation

Dataset author: J. P. Pérez Tejada (ORCID 0009-0003-9501-9918).

When you use this dataset, cite both the dataset and Ticha, the source of the images and transcriptions. If you use or adapt the code that built it, cite the repository as well.

@misc{perez_tejada_ticha_htr_lines_2026,
  author       = {P{\'e}rez Tejada, J. P.},
  title        = {Ticha manuscript lines: a line-level HTR dataset of Colonial Zapotec and Spanish manuscripts},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/maclenn77/ticha-htr-lines}},
  note         = {Derived from Ticha: a digital text explorer for Colonial Zapotec}
}

@misc{ticha,
  editor       = {Lillehaugen, Brook Danielle and Broadwell, George Aaron and Oudijk, Michel R. and
                  Allen, Laurie and Plumb, May Helena and Zarafonetis, Mike},
  title        = {Ticha: a digital text explorer for Colonial Zapotec},
  publisher    = {Haverford College Libraries},
  howpublished = {\url{https://ticha.haverford.edu}}
}

@software{perez_tejada_ticha_code_2026,
  author       = {P{\'e}rez Tejada, J. P.},
  title        = {ticha: scraping Ticha and building a line-level HTR dataset},
  year         = {2026},
  url          = {https://github.com/Maclenn77/ticha}
}