McCATMuS fine-tuned on Ticha (Colonial Zapotec and Spanish manuscripts)

A kraken text recognition model for 17th to 19th century manuscripts from Oaxaca, Mexico, in Colonial Zapotec and Spanish. It is the generic model McCATMuS fine-tuned on line images from Ticha: a digital text explorer for Colonial Zapotec.

It was trained to align transcriptions with segmented lines when building the ticha-htr-lines dataset: each line is read by the model, and the transcription is cut into the spans that match the readings. McCATMuS was never trained on Zapotec; fine-tuning brought the median character error rate of its readings on Ticha pages from 0.31 to 0.05 (measured on lines it was trained on, so optimistic) and moved 24 more pages out of manual review.

• File: ticha_ft_e06.safetensors (kraken 7 format, 16 MB; SHA-256 7c35f1554ab7c130eab65983dd8e5f394e784565417978a3d74d63c795966f1a)
• Base model: McCATMuS (McCATMuS_nfd_nofix_V1), Alix Chagué (Inria), CC BY 4.0
• Requires: kraken 7 (tested with 7.1.1)

Usage

Download the model:

from huggingface_hub import hf_hub_download

model = hf_hub_download("maclenn77/ticha-htr-mccatmus-ft", "ticha_ft_e06.safetensors")

Recognize a page with kraken (default baseline segmentation, then this model):

kraken -d cpu -i page.png page.txt segment -bl ocr -m ticha_ft_e06.safetensors

-d cpu matters on Apple Silicon: kraken otherwise picks the GPU (MPS), where it can stall. Use -d cuda:0 with an NVIDIA GPU.

Pair Ticha transcriptions with segmented lines, as for the dataset (code):

ticha-scraper lines -t texts.csv -i images.csv -d paired --model ticha_ft_e06.safetensors

The notebook notebooks/ticha_pipeline.ipynb of that repository downloads this model when FINETUNED_SOURCE = "download".

Training

Validation character accuracy after each epoch:

Limitations

• Do not use it to report results on the test split of ticha-htr-lines. Its training lines come from documents of all three splits of that dataset, test included.
• The validation lines come from the same documents as the training lines, so 86.1% overstates the accuracy on new documents and hands.
• The training transcriptions follow Ticha's conventions (editorial marks such as q[ue], (?), [roto]) and were paired automatically; a few lines may contain a word from a neighbouring line.
• It was trained on line images from good-quality colour photographs; expect lower accuracy on the faded, low-resolution notarial scans common among Ticha's untranscribed documents.

Licence and citation

The training data comes from Ticha (CC BY-NC-SA 4.0) and the base model is McCATMuS (CC BY 4.0). This model is shared under CC BY-NC-SA 4.0: non-commercial use, with attribution to the model, McCATMuS and Ticha, and derivatives under the same licence.

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

@misc{perez_tejada_ticha_mccatmus_ft_2026,
  author       = {P{\'e}rez Tejada, J. P.},
  title        = {McCATMuS fine-tuned on Ticha: a kraken model for Colonial Zapotec and Spanish manuscripts},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/maclenn77/ticha-htr-mccatmus-ft}}
}

@misc{chague_mccatmus,
  author       = {Chagu{\'e}, Alix},
  title        = {McCATMuS: Transcription model for handwritten, printed and typewritten documents from the
                  16th century to the 21st century},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.13788177}
}

@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}}
}