pplatanou/greek-letter-convnextv2-jocch

ConvNeXt-V2 Tiny LF+DSCL — Ancient Greek Character Model

This repository contains the model checkpoint and embedding artifacts associated with the interactive demonstrator accompanying the forthcoming article:

Paraskevi Platanou, Lavinia Ferretti, Isabelle Marthot-Santaniello, Giuseppe De Gregorio, Spiros Barbakos, Maria Konstantinidou, Asimina Paparrigopoulou, and John Pavlopoulos.
“Graphic Compensation in Ancient Greek Documentary Hands: A Computational Paleographic Analysis from Handwritten Character Recognition.”
ACM Journal on Computing and Cultural Heritage (JOCCH), to appear, 2026.

The repository provides the final ConvNeXt-V2 Tiny + Lacuna-based Fragmentation (LF) + Dynamically-learned Supervised Contrastive Loss (DSCL) model used in the study, together with the reference embeddings and fixed UMAP artifacts required by the interactive character-embedding demonstrator.

Citation

If you use this model, the embedding artifacts, or the interactive demonstrator, please cite:

Paraskevi Platanou, Lavinia Ferretti, Isabelle Marthot-Santaniello, Giuseppe De Gregorio, Spiros Barbakos, Maria Konstantinidou, Asimina Paparrigopoulou, and John Pavlopoulos.
Graphic Compensation in Ancient Greek Documentary Hands: A Computational Paleographic Analysis from Handwritten Character Recognition.
ACM Journal on Computing and Cultural Heritage (JOCCH), to appear, 2026.

@article{platanou2026graphic, author = {Platanou, Paraskevi and Ferretti, Lavinia and Marthot-Santaniello, Isabelle and De Gregorio, Giuseppe and Barbakos, Spiros and Konstantinidou, Maria and Paparrigopoulou, Asimina and Pavlopoulos, John}, title = {Graphic Compensation in Ancient Greek Documentary Hands: A Computational Paleographic Analysis from Handwritten Character Recognition}, journal = {ACM Journal on Computing and Cultural Heritage}, year = {2026}, note = {To appear} }

Model

• Backbone: ConvNeXt-V2 Tiny
• Classes: 24 Ancient Greek letter classes
• Input: grayscale single-character crops
• Evaluation input size: 64 × 64
• Training strategy: Lacuna-based Fragmentation (LF) + Dynamically-learned Supervised Contrastive Loss (DSCL)
• Final test accuracy: 0.8643872454859777
• Final macro-F1: 0.8599418651059835

Repository contents

best_convnextv2_tiny_lf_dscl.pth
artifacts/
  demo_config.json
  hellchar_reference_embeddings.npy
  hellchar_reference_metadata.csv
  hellchar_reference_umap.npy
  umap_reducer.joblib

The UMAP artifacts support interactive visualization. Quantitative nearest-neighbour retrieval is performed in the original normalized ConvNeXt embedding space using cosine similarity.

Associated resources

• Source code and datasets:
  https://github.com/ipavlopoulos/diachronic-greek-letterforms

• Interactive demonstrator:
  https://diachronic-greek-letterforms.streamlit.app

• Publication:
  Graphic Compensation in Ancient Greek Documentary Hands: A Computational Paleographic Analysis from Handwritten Character Recognition, ACM Journal on Computing and Cultural Heritage (JOCCH), to appear, 2026.

Intended use

Research and demonstration on historical handwritten Ancient Greek single-character crops. The model is not intended for modern fonts, full text lines, or unrelated images.

License / data provenance

he model and accompanying reproducibility artifacts are released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license, consistent with the accompanying diachronic-greek-letterforms repository.

The model was trained on Hell-Char, a curated character-level subset of Hell-Date consisting of Ancient Greek handwritten letter crops. Hell-Char and the associated data resources are documented in the accompanying repository. The underlying datasets build on securely dated historical papyrological and manuscript material; users should also cite the originating datasets and resources where applicable.

The model checkpoint and derived embedding artifacts are provided for research, reproducibility, and scholarly use. Use and redistribution are subject to the attribution requirements of the CC BY 4.0 license. Users of the model or derived artifacts should cite the associated publication and source repository.