mbruton/LegibilisCorpus-models

LegibilisCorpus Models

Trained HTR (Handwritten Text Recognition) models from "Language Similarity and Cross-Lingual Transfer in Historical HTR: Evidence from Swedish, Norwegian, and Medieval Latin", presented at the 8th International Workshop on Historical Document Imaging and Processing (HIP'26), held in conjunction with ICDAR 2026.

• Paper code / experiment scripts: github.com/mbruton0426/historical-htr-LegibilisCorpus
• Dataset: huggingface.co/datasets/mbruton/LegibilisCorpus

Overview

These checkpoints were trained as part of a fully crossed cross-lingual transfer study across three historical line-level HTR datasets — Riksarkivet (Swedish), NorHand v3 (Norwegian Bokmål), and HOME-Alcar (Medieval Latin / Old French) — using two HTR pipelines, Loghi and Kraken.

Repository Contents

.
├── kraken/    # Trained Kraken model checkpoints (.zip)
└── loghi/     # Trained Loghi model checkpoints (.zip)

Each pipeline folder contains the full set of models trained for the paper's experiments:

• Base models — one per source dataset (Riksarkivet, NorHand, HOME-Alcar), trained from scratch to convergence.
• Fine-tuned models — each base model fine-tuned separately on each of the remaining two datasets (6 fine-tuned variants per pipeline).
• Low-resource models — models trained on reduced data (5,000 and 15,000 lines), both per-dataset and on pooled multilingual data.

30 of 32 models per pipeline converged during training; for non-converged models, the checkpoint with the lowest loss and highest accuracy was retained (see paper §3.3 and §5.4 for details).

Intended Use

These models are intended for research and reproducibility purposes — to support replication of the paper's zero-shot transfer, fine-tuning, and low-resource evaluation results, and as a reference point for further work on multilingual and cross-lingual historical HTR.

They are trained on 19th–20th century Swedish and Norwegian administrative/personal documents and 12th–14th century Latin/Old French cartularies, and should not be assumed to generalize well outside these domains, time periods, or scripts without further fine-tuning.

Usage

Models are provided as pipeline-native checkpoints, compatible with:

• Loghi
• Kraken

Download and unzip the relevant checkpoint, then follow the respective pipeline's documentation for loading a pretrained model for inference or further fine-tuning. Sample training/evaluation commands used to produce these models are available in the GitHub repository's bash/ folder.

Citation

A full citation (including proceedings volume and page numbers) will be added once available following presentation at HIP'26.

@inproceedings{bruton2026crosslingual,
  author    = {Bruton, Micaella and Tudor, Crina and Sinnaeve, Wout and Yousuf, Oreen and Rirdance, Signe and Heil, Raphaela and Megyesi, Be{\'a}ta},
  title     = {Language Similarity and Cross-Lingual Transfer in Historical {HTR}: Evidence from {S}wedish, {N}orwegian, and {M}edieval {L}atin},
  booktitle = {Proceedings of the 8th International Workshop on Historical Document Imaging and Processing (HIP'26)},
  year      = {2026},
  note      = {To appear},
  address   = {ICDAR 2026}
}

Acknowledgments

This work is financed in part by the Swedish Graduate School of Digital Philology through the Swedish Research Council (grant 2022-06343), and supported by Riksbankens Jubileumsfond, grant M24-0028 (Echoes of History: Analysis and Decipherment of Historical Writings, DESCRYPT). Computations were enabled by resources provided by the National Academic Infrastructure for Supercomputing in Sweden (NAISS), partially funded by the Swedish Research Council (grant 2022-06725).

License

Released under the MIT License unless noted otherwise.