small-models-for-glam/kraken-blla-base

BLLA base model (kraken default segmentation model)

Hugging Face Hub mirror

This repository mirrors the canonical Zenodo release by Benjamin Kiessling, ALMAnaCH, Inria Paris. The files are byte-identical to the Zenodo release and have not been modified by Small Models for GLAM.

As of kraken 7.1.1 (the latest PyPI release on 2026-10-24) the package still bundles the previous cBAD-2019-only BLLA model. Download this repository to use the new default model now.

Use from the Hugging Face Hub

pip install "kraken>=7.1.0" huggingface_hub

hf download small-models-for-glam/kraken-blla-base blla.mlmodel --local-dir ./kraken-blla-base

kraken -i image.png output.txt \
    segment -bl -i ./kraken-blla-base/blla.mlmodel \
    ocr -m <recognition model>

For a recognition model, see the Kraken PP-OCRv6 collection. The training splits (train.lst, val.lst, test.lst) are mirrored as well.

This is the default segmentation model shipped with the kraken ATR engine. It replaces the previous base model trained on cBAD 2019 alone and has been retrained from scratch on a considerably larger and more diverse corpus of ~11200 pages aggregated from six datasets.

Uses

The model performs reasonably well on most non-fragmentary handwritten and machine-printed document pages of moderate complexity. If out of the box performance is unsatisfactory or more fine-grained segmentation taxonomies are desired it can also serve as a good foundation model for fine-tuning.

Training Data

The training corpus aggregates six publicly available datasets. Pages were split 11,271 / 1,000 / 1,000 into training, validation, and test sets at random.

Dataset	Source	Train	Val	Test	Total
OCR-D-LAYOUT-GT-DTA	10.5281/zenodo.19348674	3,741	346	326	4,413
Burchards Dekret Digital	10.5281/zenodo.10882817	3,035	246	257	3,538
cBAD 2019	10.5281/zenodo.3568023	2,582	209	230	3,021
Muharaf	10.5281/zenodo.11492215	1,011	101	104	1,216
CLLG	gitlab.inria.fr/almanach/cllg	460	47	40	547
Baybars	gitlab.huma-num.fr/lipa/baybars	442	51	43	536
Total		11,271	1,000	1,000	13,271

The exact splits are deposited alongside the model as train.lst, val.lst, and test.lst. Each line is the path of a PAGE or ALTO file relative to the root of the combined corpus. The datasets are freely licensed and can be obtained from the sources above.

Region and line types in the training dataset have been merged into single classes.

Biases and Limitations

The goal of this training run was to take advantage of the superior accuracy of models trained with the recently implemented hybrid CE+Dice loss but also to improve accuracy on machine-printed documents with close writing, images taken with handheld devices, and non-Latin material. Our tests indicate that this was generally successful, nevertheless the model still exhibits some of the known weaknesses.

The training corpus consists exclusively of Latin, Arabic, and Greek script documents with relatively simple layouts. Accuracy on scripts and layouts dissimilar to those three is therefore expected to be uneven. Some features of pages where segmentation accuracy is frequently unsatisfactory:

• Tilted and vertical writing
• Writing embedded in decoration

An idiosyncracy of the model is that no matter the position of the calligraphic/typographic baseline (top-, center-, or baseline) a baseline will be detected. Hebrew script lines will therefore be detected with base- not toplines. This is generally not an issue for recognition models that have been trained with the same line position.

How to Get Started with the Model

Install the kraken ATR package following the instruction here. A quick start explaining the use of the segmenter can also be found there.

Training

The model was trained from scratch with kraken's default BLLA architecture for 20 epochs on the 11,271 page training split, with data augmentation enabled. The optimizer was AdamW at a learning rate of 2e-4 with a weight decay of 1e-5, gradient clipping at 1.0, and a cosine schedule with 2,000 warmup steps decaying to a minimum learning rate of 2e-5.

Evaluation

Evaluated with ketos segtest on the held-out 1,000 page test split. The line evaluation uses kraken's adaptation of the cBAD baseline evaluation metric:

Precision	Recall	F1
0.964	0.947	0.950

Pixel accuracy and IOU for the text region:

Pixel accuracy	IOU
0.965	0.929

Acknowledgements

Training of this model was funded by the European Union under Grant Agreement No.~101132163 (ATRIUM). Views and opinions expressed are those of the authors only and do not necessarily reflect those of the European Union.

Citation

If you use this model, please cite kraken and if possible credit the dataset providers linked in the Training Data table above.

BibTeX:

@inproceedings{10.1007/978-3-032-04624-6_26,
  author = {Kiessling, Benjamin},
  title = {Version 5 of the Kraken ATR Engine for the Humanities},
  year = {2025},
  isbn = {978-3-032-04623-9},
  publisher = {Springer-Verlag},
  address = {Berlin, Heidelberg},
  doi = {10.1007/978-3-032-04624-6_26},
  booktitle = {Document Analysis and Recognition – ICDAR 2025: 19th International Conference, Wuhan, China, September 16–21, 2025, Proceedings, Part III},
  pages = {443–458},
  numpages = {16},
  location = {Wuhan, China}
}