ENC-PSL/Medusa0.1Line-9B

MEDUSA 0.1 — Medieval European Documents Unified System for Automated text recognition

MEDUSA is a family of Vision Language Models (VLMs) fine-tuned for multilingual medieval handwritten text recognition (HTR) at the line level. These models were developed at the École nationale des chartes – PSL and submitted to the ICDAR 2026 Competition on Multilingual Medieval Handwritten Text Recognition (CMMHWR).

This repository contains two model variants:

Model	Base	Version	Notes
MEDUSA-4B-0.1	Qwen3.5-4B	0.1	Model used at competition submission
MEDUSA-9B-0.1	Qwen3.5-9B	0.1	Best model at competition submission

Note on versioning. The 0.1 variants correspond exactly to the models evaluated in the ICDAR 2026 competition.

---

System report

For full details on data, training procedure, and results, see the accompanying system report.

---

Languages and scripts

MEDUSA was trained on a corpus of over 640,000 lines spanning more than twenty repositories, covering the following language families and scripts:

• Romance / Latin: Old French (fro), Occitan (pro), Old Italian (ita), Old Spanish (osp), Catalan (cat), Old Portuguese (opor), Navarrese (nav), Latin (lat), Venetian (vec), Galician (glg)
• Germanic: Middle High German (gmh), Middle Low German (gml), Old Icelandic (ice), Middle English (enm), Middle Dutch (dum), Old English (ang), Old Norwegian (non), Swedish (swe)
• Celtic: Welsh (wlm), Old Irish (gle)
• Slavic: Old Czech (cze), Old Polish (pol)

Mnuscripts dated roughly from the 9th to the 15th century.

---

Results (ICDAR 2026 CMMHWR)

Unweighted average CER (%) and WER (%) on internal and official competition test sets. Lower is better.

Model	Internal CER	Internal WER	Task 1 CER	Task 2 CER	Task 3 CER
kraken-CATMuS 1.6.0 (baseline)	17.3	53.5	9.29	7.91	25.9
MEDUSA-4B 0.1	14.7	44.5	8.15	5.60	12.0
MEDUSA-9B 0.1	13.2	42.6	8.03	5.24	10.8

---

Training data

MEDUSA 0.1 is trained on two tiers of image–text data:

• Gold — paired image–text data following heterogeneous transcription conventions, used for visual adaptation across manuscript styles and editorial traditions.
• Platinum — image–text pairs aligned with the CATMuS diplomatic transcription guidelines, used for final specialization toward the target task.

The total training pool amounts to approximately 645,000 line-level image–text pairs.

Gold datasets

Dataset	Language	Level	Lines
Original data	Multilingual	Page	18,352
COMETA [1]	Occitan	Page	118,105
Torino L-II-14 [2]	Old French	Page	36,823
Tridis [3]	Multilingual	Line	166,784
FROC-MSS [4]	Old French / Occitan	Page	3,636
iForal [5]	Latin / Old Portuguese	Page	8,009
DISTINGUO [6]	Latin	Page	15,190
AMSMB [7]	Latin / Catalan	Page	3,369
HTR-School-Vienna-2025 [8]	Latin	Page	7,477
Paris Bible Project [9]	Latin	Page	1,606
St-Victor (M. Vernet) [10]	Latin	Page	10,736
Wien ÖNB Cod. 2160 f. 164-184 [11]	Latin	Page	2,681
Bifrost [12]	Old Norse	Page	873
Klosterneuburg [13]	Middle High German	Page	4,758
Faithful transcriptions [14]	Multilingual	Page	8,001
StABS Ratsbücher [15]	Middle High German	Page	8,371
Inzigkofen [16]	Middle High German	Page	8,321
Total (Gold)			423,092

Platinum datasets

Dataset	Language	Level	Lines
Original data	Multilingual	Page	10,506
CMMHWR dataset [17]	Multilingual	Page	149,741
CATMuS Medieval [18]	Middle Dutch, Old English	Line	47,084
GATMUZA [19]	Occitan	Page	2,117
TranscriboQuest 2025 [20]	Multilingual	Page	1,278
HTR-School-Vienna-02 [21]	Old Czech	Page	1,336
Padeřov-Bible [22]	Old Czech	Page	7,177
2024–medieval-czech [23]	Old Czech	Page	2,748
Total (Platinum)			221,987

References

[1] Wiedner, M. COMETA : Corpus de l'occitan médiéval comparatif et annoté. https://zenodo.org/records/15300719

[2] Camps, J.-B., O'Connor, P. Torino_L-II-14: HTR Training Dataset for the manuscript Turin, Biblioteca nazionale universitaria, MS L. II. 14 (2024). https://github.com/RESCAPE-Biblissima/Torino_L-II-14

[3] Torres, S. Tridis (revision e8d811f) (2025). https://doi.org/10.57967/hf/5001

[4] Camps, J.-B. FROC-MSS: Old French and Old Occitan Medieval Manuscripts HTR Data and Models (2018). https://github.com/Jean-Baptiste-Camps/FROC-MSS

[5] Projet iForal. iForal Dataset: Medieval Portuguese Manuscripts HTR Data. https://github.com/Arch-W/iForal-Dataset

[6] Burghart, M., Yatsyk, S. DISTINGUO: Ground truth for handwritten text recognition (HTR) on collections of distinctions (late 13th to late 15th century) (2024). https://doi.org/10.34847/NKL.48AD8B8D

[7] Coll Ardanuy, M., Cuadrada, C., Sarobe, R. A Dataset for Handwritten Text Recognition in Medieval Notarial Charters Written on Parchment (2025). https://dataverse.bsc.es/dataset.xhtml?persistentId=perma:BSC/0VB0MC

[8] Odstrčilík, J. et al. HTR Winter School in Vienna 2025 – Late Medieval Latin Group: Ground Truth Dataset for Late-Medieval Latin Scripts (2025).

[9] Wrisley, D., The Paris Bible Project, Gueville, E. parisbible/ground_truth: Ground truth v1.0.0 for the Paris Bible Project (Feb 2023). https://doi.org/10.5281/zenodo.7653691

[10] Vernet, M. Saint Victor MS dataset (abbreviated and expanded ALTO) (Jan 2023). https://doi.org/10.5281/zenodo.7510410

[11] Attwood et al. Wien ÖNB cod. 2160 f. 164-184 ground truth from HTR Winter School 2022 (Dec 2022). https://doi.org/10.5281/zenodo.7537204

[12] Kapitan, K.A., Vidal-Gorène, C. Crossing the Bifrost: Towards an open access FAIR HTR model for Old Norse manuscripts (May 2025). https://doi.org/10.5281/zenodo.15366896

[13] Berger, M. et al. Klosterneuburg, Stiftsbibl., Cod. 48 – Ground Truth: Initial release (Dec 2022). https://doi.org/10.5281/zenodo.7466928

[14] Eichenberger, N., Suwelack, H. Faithful Transcriptions Data Set: TEI/XML-encoded transcriptions of medieval theological manuscripts (Oct 2021). https://doi.org/10.5281/zenodo.5582483

[15] Hodel, T., Schoch, D., Dängeli, P. Handwritten text recognition ground truth set: StABS Ratsbücher O10, Urfehdenbuch X (Aug 2021). https://doi.org/10.5281/zenodo.5153263

[16] Eichenberger, N. Transcriptions from medieval manuscripts related to the Augustinian canonesses in Inzigkofen (Dec 2025). https://doi.org/10.5281/zenodo.17978574

[17] Clérice, T., Kiessling, B. ICDAR 2026 Competition on Multilingual Medieval Handwriting Recognition (Jan 2026). https://doi.org/10.5281/zenodo.18270331

[18] Clérice, T. et al. CATMuS Medieval: A multilingual large-scale cross-century dataset in Latin script for handwritten text recognition and beyond (Feb 2024). https://inria.hal.science/hal-04453952

[19] Camps, J.-B. Lo GAT MUZA: CATMuS-conformant HTR Ground-Truth Data for Medieval Occitan (2026). https://github.com/LostMa-ERC/gatmuza

[20] McDonough, C. et al. TranscriboQuest 2025 Medieval Vernacular Religious Texts (Sep 2025). https://doi.org/10.5281/zenodo.17062963

[21] Veličkaitė, V. et al. HTR Winter School 2025 – Medieval Czech – Biblioteka Jagiellonska BJ Rkp 441 IV (Dec 2025).

[22] Michalcová, A. et al. Padeřov-Bible-handwriting-ground-truth: Initial release (Dec 2022). https://doi.org/10.5281/zenodo.7467034

[23] Plechatý, M. et al. HTR Winter School 2024 – Medieval Czech – Prague Bible (1488) (Dec 2024).

---

Intended use

These models are designed for line-level HTR on pre-segmented medieval manuscript images. They are not page-level OCR systems: they expect a cropped image of a single text line as input and return the transcription of that line.

The models target CATMuS transcription guidelines, which govern abbreviation expansion, allograph normalisation, and the character set used.

---

Usage with DocWorkflow

The recommended way to use MEDUSA is via DocWorkflow, the document analysis framework developed at the École nationale des chartes. DocWorkflow handles ALTO XML input/output, line image extraction, batching, CATMuS post-processing, and scoring in a unified pipeline.

Installation

git clone https://github.com/TheoMoins/DocWorkflow
cd docworkflow
pip install -e .

Configuration file

Create a YAML config file (e.g., medusa_inference.yml):

run_name: "Medusa0.1Line-9B"
output_dir: "results"
device: "cuda"
save_image: true

data:
  test: "path/to/your/alto/data"   # directory with ALTO XML + image pairs

tasks:
  htr:
    type: VLMLineHTR
    config:
      use_metadata: true
      model_name: 'outputs/Medusa0.1Line-9B'
      device_map: "auto"
      max_new_tokens: 128
      line_batch_size: 8
      max_pixels: 401408

      prompt: >
        Transcribe the handwritten text in this line image.
        Keep abbreviations as written, do not expand them.
        Modernize word segmentation (split or join words following modern usage).
        Use only u and i, never v and j, regardless of the original or modern usage.
        Do not record allographic variants, use standard letter forms.
        Output ONLY the transcription.

Tip. If your dataset follows dataset-specific transcription conventions, you can provide a conventions.yml file alongside your ALTO data. DocWorkflow will automatically inject the conventions into the prompt via a {conventions} placeholder.

Running inference

docworkflow -c Medusa0.1Line-9B.yml predict -t htr -d test

Direct usage with transformers

The models can also be used directly outside of DocWorkflow, though the CATMuS post-processing step will need to be applied manually if desired.

from transformers import AutoProcessor, AutoModelForImageTextToText
from PIL import Image

model_id = "ENC-PSL/Medusa0.1Line-9B"

processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(model_id, device_map="auto")

image = Image.open("path/to/line_image.jpg").convert("RGB")

prompt = "Transcribe the handwritten text in this line image.\nOutput ONLY the transcription."

messages = [
    {
        "role": "user",
        "content": [
            {"type": "text", "text": prompt},
            {"type": "image", "image": image},
        ],
    }
]

inputs = processor.apply_chat_template(
    [messages],
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    enable_thinking=False,
    return_tensors="pt",
).to(model.device)

with torch.no_grad():
    generated_ids = model.generate(**inputs, max_new_tokens=128, do_sample=False)

trimmed = generated_ids[0][inputs["input_ids"].shape[1]:]
transcription = processor.decode(trimmed, skip_special_tokens=True).strip()
print(transcription)

Important. The model is optimised for the prompt given above. Results may degrade if the prompt is modified. The model outputs raw text; to enforce CATMuS compliance (character whitelist, allograph normalisation), apply the post-processing step provided in DocWorkflow (src/tasks/htr/postprocessing.py).

---

CATMuS conventions

Transcriptions follow the CATMuS guidelines.

---

Citation

If you use MEDUSA in your research, please cite:

@unpublished{moins:hal-05600991,
  TITLE = {{MEDUSA 0.1: Medieval European Documents Unified System for Automated text recognition System Report for the ICDAR 2026 Competition on Multilingual Medieval Handwritten Text Recognition}},
  AUTHOR = {Moins, Th{\'e}o and Cafiero, Florian and Camps, Jean-Baptiste and Conte, Lilla and Guidi, Emilie and Hensley, Brenna and Kapitan, Katarzyna and Macedo, Carolina and Peratello, Paola and Vermaas, Cecile and Vidal-Gor{\`e}ne, Chahan},
  URL = {https://enc.hal.science/hal-05600991},
  NOTE = {working paper or preprint},
  YEAR = {2026},
  MONTH = Apr,
  PDF = {https://enc.hal.science/hal-05600991v1/file/MEDUSA__MEDieval_Universal_Script_Analysis-6.pdf},
  HAL_ID = {hal-05600991},
  HAL_VERSION = {v1},
}

---

Funding

Funded by the European Union (ERC, LostMA, 101117408). Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Council. Neither the European Union nor the granting authority can be held responsible for them.

This work has received support under the Major Research Program of PSL Research University ``CultureLab'' launched by PSL Research University and implemented by ANR with the references ANR-10-IDEX-0001.

Ce travail a bénéficié d'une aide de l’État gérée par l'Agence Nationale de la Recherche au titre de France 2030 portant la référence « ANR-23-IACL-0008»).

Biblissima+ bénéficie d’une aide de l'Etat gérée par l'ANR au titre du Programme d’investissements d’avenir intégré à France 2030, portant la référence ANR-21-ESRE-0005.

This work was granted access to the HPC resources of IDRIS under the allocation 2026-AD011015914R1 made by GENCI.