Create README.md

README.md

@@ -0,0 +1,310 @@
+---
+license: cc-by-4.0
+datasets:
+- CATMuS/medieval
+- magistermilitum/Tridis
+metrics:
+- cer
+- wer
+pipeline_tag: image-to-text
+tags:
+- htr
+- handwritten-text-recognition
+- manuscripts
+- medieval
+- vision-language-model
+- catmus
+- qwen
+---
+
+# MEDUSA 0.2 — 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](https://www.chartes.psl.eu/).
+
+This repository contains the `0.2` variants of MEDUSA, which extend the [0.1 release](https://huggingface.co/ENC-PSL/Medusa0.1Line-9B) with an explicit 
+**Silver pre-training stage** on synthetic line images rendered from text-only historical corpora.
+
+> **Note on versioning.** The `0.1` variants correspond to the models submitted to the ICDAR 2026 competition, trained exclusively on Gold and Platinum image–text pairs.
+> The `0.2` variants add a prior Silver stage using 500,000 synthetic lines, improving coverage of under-resourced languages.
+
+---
+
+## System report
+
+For full details on data, training procedure, and results, see the [accompanying system report](https://enc.hal.science/hal-05600991).
+
+---
+
+## Languages and scripts
+
+MEDUSA was trained on a corpus of over **640k lines + 500k generated 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`)
+
+Manuscripts dated roughly from the 9th to the 15th century.
+
+---
+
+## Training data and curriculum
+
+MEDUSA 0.2 is trained in three successive stages:
+
+### Stage 1 — Silver (synthetic)
+
+The Silver stage uses **text-only historical corpora** to generate **500,000 synthetic line images**, providing lexical and script coverage for languages that are absent or scarce in the image–text tiers. Text is sampled from over 20 million words across medieval languages, rendered onto real manuscript background patches with stochastic ink, scan, and document degradation effects.
+
+The synthetic lines are available as a separate dataset: [ENC-PSL/MEDUSA-synthetic-lines](https://huggingface.co/datasets/ENC-PSL/MEDUSA-synthetic-lines).
+
+The underlying text-only corpora are listed below.
+
+| Dataset | Language | Reference |
+|---|---|---|
+| Helsinki Corpus | Old English / Middle English | [[1]](#ref-helsinki) |
+| Floris and Blancheflour | Old English | [[2]](#ref-floris) |
+| Diakorp | Old Czech | [[3]](#ref-diakorp) |
+| Hadewijch | Middle Dutch | [[4]](#ref-hadewijch) |
+| CELT | Old Irish / Middle Irish | [[5]](#ref-celt) |
+| Referenzkorpus Altdeutsch | Old High German | [[6]](#ref-altdeutsch) |
+| Referenzkorpus Mittelhochdeutsch | Middle High German | [[7]](#ref-mhd) |
+| Referenzkorpus Mittelniederdeutsch | Middle Low German | [[8]](#ref-mnd) |
+| CTA | Old / Middle Portuguese | [[9]](#ref-cta) |
+| CIPM | Old / Middle Portuguese | [[10]](#ref-cipm) |
+| CODEA | Old Castilian | [[11]](#ref-codea) |
+| OSTA | Old Castilian | [[12]](#ref-osta) |
+| Biblia Medieval | Old Castilian | [[13]](#ref-biblia) |
+| Milione | Venetian | [[14]](#ref-milione) |
+| BFM | Old / Middle French | [[15]](#ref-bfm) |
+| Chrétien de Troyes | Old / Middle French | [[16]](#ref-chretien) |
+| Geste | Old / Middle French | [[17]](#ref-geste) |
+| Otinel | Old / Middle French | [[18]](#ref-otinel) |
+| Menota | Old Norse / Icelandic | [[19]](#ref-menota) |
+| EAE | Old Norse / Icelandic | [[20]](#ref-eae) |
+| Flores och Blanzeflor | Old Swedish | [[21]](#ref-flores) |
+| Italian Paleography | Old Italian | [[22]](#ref-italpal) |
+| CLTK Latin Library | Latin | [[23]](#ref-cltk) |
+| Medieval Latin | Latin | [[24]](#ref-medlat) |
+| Kochanowski | Old Polish | [[25]](#ref-kochanowski) |
+
+### Stage 2 — GoldMEDUSA 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.
+
+Paired image–text data following heterogeneous transcription conventions, used for visual adaptation across manuscript styles and editorial traditions (~423,000 lines). See the [MEDUSA 0.1 model card](https://huggingface.co/ENC-PSL/Medusa0.1Line-9B) for the full dataset table.
+
+### Stage 3 — Platinum
+
+Image–text pairs aligned with the [CATMuS](https://catmus-guidelines.github.io/) diplomatic transcription guidelines, used for final specialization toward the target task (~222,000 lines). See the [MEDUSA 0.1 model card](https://huggingface.co/ENC-PSL/Medusa0.1Line-9B) for the full dataset table.
+
+---
+
+### Silver-stage references
+
+<a id="ref-helsinki"></a>[1] *Helsinki Corpus TEI XML Edition* (2011). Designed by Alpo Honkapohja et al. Based on The Helsinki Corpus of English Texts (1991).
+
+<a id="ref-floris"></a>[2] Draschner, M., Edlich-Muth, M. *Raw text edition of the Middle English 'Floris and Blancheflour' in Edinburgh, National Library of Scotland, MS Advocates 19.2.1* (Jan 2026). https://doi.org/10.5281/zenodo.18244892
+
+<a id="ref-diakorp"></a>[3] Kučera, K., Řehořková, A., Stluka, M. *Diakorp v6: Diachronic Corpus of Czech*. LINDAT/CLARIAH-CZ (2015). http://hdl.handle.net/11234/1-5413
+
+<a id="ref-hadewijch"></a>[4] Haverals, W., Kestemont, M. *From exemplar to copy: the scribal appropriation of a Hadewijch manuscript computationally explored*. JDMDH 23. https://doi.org/10.46298/jdmdh.10206
+
+<a id="ref-celt"></a>[5] Ó Corráin, D. et al. *CELT: Corpus of Electronic Texts* (1997). http://www.ucc.ie/celt
+
+<a id="ref-altdeutsch"></a>[6] Donhauser, K. et al. *Referenzkorpus Altdeutsch (750–1050)* (2022), version 1.2. https://www.deutschdiachrondigital.de/rea/
+
+<a id="ref-mhd"></a>[7] Roussel, A. et al. *Referenzkorpus Mittelhochdeutsch (1050–1350)* (2024), version 2.1. https://www.linguistics.ruhr-uni-bochum.de/rem/
+
+<a id="ref-mnd"></a>[8] ReN-Team. *Reference Corpus Middle Low German/Low Rhenish (1200–1650)* (2021), version 1.1. https://doi.org/10.25592/uhhfdm.9195
+
+<a id="ref-cta"></a>[9] Sobral, C., Cardeira, E. *Corpus de Textos Antigos (CTA)*. http://teitok.clul.ul.pt/teitok/cta/
+
+<a id="ref-cipm"></a>[10] Xavier, M.F. *O CIPM — Corpus Informatizado do Português Medieval*. In: Kabatek, J. (ed.), De Gruyter (2016). https://doi.org/10.1515/9783110462357-007
+
+<a id="ref-codea"></a>[11] GITHE. *CODEA+ 2022. Corpus de documentos españoles anteriores a 1900* (2022). https://doi.org/10.37536/CODEA.2015
+
+<a id="ref-osta"></a>[12] *Old Spanish Textual Archive (OSTA)* (2026). https://github.com/hispanicseminary/OSTA
+
+<a id="ref-biblia"></a>[13] Enrique-Arias, A., Pueyo Mena, F.J. *Biblia Medieval* (2008). https://bibliamedieval.es
+
+<a id="ref-milione"></a>[14] Burgio, E. et al. *Dei Viaggi di Messer Marco Polo [...]: Edizione digitale* (2015). https://risorse-esterne.edizionicafoscari.it/
+
+<a id="ref-bfm"></a>[15] Guillot, C., Heiden, S., Lavrentiev, A. *Base de français médiéval*. Diachroniques 7, 168–184 (2018). https://shs.hal.science/halshs-01809581
+
+<a id="ref-chretien"></a>[16] Kunstmann, P., Martineau, F. *Chrétien de Troyes sur le web* (2000). https://doi.org/10.16995/dscn.176
+
+<a id="ref-geste"></a>[17] Camps, J.-B. et al. *Geste: un corpus de chansons de geste, 2016–...* (Apr 2019). https://doi.org/10.5281/zenodo.2630574
+
+<a id="ref-otinel"></a>[18] Camps, J.-B. *La Chanson d'Otinel: édition complète du corpus manuscrit* (Dec 2017). https://doi.org/10.5281/zenodo.1116736
+
+<a id="ref-menota"></a>[19] *Medieval Nordic Text Archive (Menota)*. https://clarino.uib.no/menota/home. Founded 2001, part of CLARIN since 2016.
+
+<a id="ref-eae"></a>[20] The Arnamagnæan Institute, Copenhagen. *Editiones Arnamagnæanæ Electronicæ (EAE)*. https://eae.ku.dk/
+
+<a id="ref-flores"></a>[21] Worrall, E., Edlich-Muth, M. *Raw text edition of the Old Swedish "Flores och Blanzeflor"* (Sep 2025). https://doi.org/10.5281/zenodo.17093371
+
+<a id="ref-italpal"></a>[22] Magni, I., Markey, L., Signorini, M. *Italian Paleography* (2019). https://italian.newberry.t-pen.org/
+
+<a id="ref-cltk"></a>[23] *Classical Language Toolkit (CLTK): lat_text_latin_library*. https://github.com/cltk/lat_text_latin_library. Texts in the public domain.
+
+<a id="ref-medlat"></a>[24] Corbara, S. et al. *Two datasets for the computational authorship analysis of medieval Latin texts* (Jun 2020). https://doi.org/10.5281/zenodo.4298503
+
+<a id="ref-kochanowski"></a>[25] Kochanowski, J. *Works of Jan Kochanowski*. Wikiźródła (Polish Wikisource). https://pl.wikisource.org/wiki/Autor:Jan_Kochanowski
+
+---
+
+## 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](https://github.com/Enceintes-Chartes/docworkflow)**, 
+the document analysis framework developed at the École nationale des chartes.
+
+### Installation
+
+```bash
+git clone https://github.com/TheoMoins/DocWorkflow
+cd docworkflow
+pip install -e .
+```
+
+### Configuration file
+
+```yaml
+run_name: "Medusa0.2Line-9B"
+output_dir: "results"
+device: "cuda"
+save_image: true
+
+data:
+  test: "path/to/your/alto/data"
+
+tasks:
+  htr:
+    type: VLMLineHTR
+    config:
+      use_metadata: true
+      model_name: 'outputs/Medusa0.2Line-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.
+```
+
+### Running inference
+
+```bash
+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.
+
+```python
+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](https://catmus-guidelines.github.io/). 
+
+---
+
+## Citation
+
+If you use MEDUSA in your research, please cite:
+
+```bibtex
+@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 = {v3},
+}
+```
+
+---
+
+## 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.