Hell-Char: Greek Handwriting Recognition

• Hell-Char is a restructured subset of the Hell-Date dataset designed for character-level learning in ancient Greek handwriting, capturing diachronic variations and manuscript degradation.
• It comprises 13,046 annotated character images from 157 papyri, reflecting imbalanced class distributions and diverse preservation qualities typical of historical documents.
• The dataset supports robust embedding learning through methods like DSCL and lacuna-driven augmentation, enabling detailed paleographic analysis and temporal generalization testing.

Hell-Char is a restructured subset of the Hell-Date papyrus dataset used for character recognition and diachronic representation learning in ancient Greek handwriting. It was introduced as a curated training set spanning the 3rd–1st centuries BCE, alongside the evaluation sets PaLit-Char and Med-Char, in a study of representations that remain robust across centuries of variation in handwriting. Within that framework, Hell-Char is the training and benchmarking base for learning embeddings that are robust to scarce data, manuscript corruption, and fine-grained inter-letter structure (Pavlopoulos et al., 23 Jun 2026).

1. Provenance and chronological scope

Hell-Char derives from Hell-Date, a papyrus dataset containing 194 images from 157 papyri, all Greek, dated between 310 BCE and 3 BCE. Of those images, 171 are annotated at character level. Hell-Char filters and reorganizes those character annotations into a corpus specifically designed for character-level learning tasks (Pavlopoulos et al., 23 Jun 2026).

Its chronological coverage is the last three centuries BCE. The paper characterizes this interval as a formative stage in Greek handwriting, when epigraphic letterforms are increasingly influenced by cursive writing, but before the later major shift to Byzantine minuscule. In diachronic terms, Hell-Char therefore occupies the Hellenistic segment of the full experimental design and anchors later temporal generalization tests.

The dataset is not presented as a generic OCR benchmark. Its intended role is narrower and more technical: to supply historically grounded, character-level training data for representation learning under temporal drift, data scarcity, and physical degradation.

2. Curation protocol and label structure

The source annotations in Hell-Date include 29 classes: the 24 standard Greek letters, the 3 archaic numeral letters Stigma, Qoppa, Sampi, a general symbol class, and an unknown class for uncertain signs, though it was empty. In Hell-Char, archaic numerals, symbols, and unknown signs were merged into a single “Unknown” category, which was then ignored in later analysis (Pavlopoulos et al., 23 Jun 2026).

The curation pipeline applies two additional filters. Only characters with BT1 preservation quality were kept, and, to reduce imbalance, the authors kept at most five instances per character class per papyrus, sampled randomly. After these steps, the final Hell-Char subset contains 13,046 character images from 157 papyri.

These curation choices are methodologically important. They preserve the historical heterogeneity of the material while constraining annotation noise and extreme within-document redundancy. A plausible implication is that the resulting corpus is intended to balance fidelity to manuscript conditions against the statistical requirements of supervised learning.

3. Data characteristics and benchmark difficulty

Hell-Char contains handwritten ancient Greek letter images with substantial variation in stroke shape, scale, slant, cursivity, preservation quality, and manuscript noise and degradation. The paper emphasizes that the dataset remains imbalanced, reflecting Greek letter frequency: high-frequency letters like Alpha dominate, while rare letters such as Psi, Zeta, Xi occur much less often.

This imbalance is treated as a feature of the benchmark rather than an artifact to be fully normalized away. The dataset is described as a hard and realistic benchmark rather than a clean OCR-style dataset. That characterization is consistent with its use in representation learning settings where the objective is not merely closed-set recognition on homogeneous scans, but robust embedding formation under partial visibility, fragmentary ink, and historically variable glyph morphology.

The work also argues that generic natural-image features do not transfer well to historical Greek handwriting. In that sense, Hell-Char is designed to expose failures of off-the-shelf visual representations and to motivate domain-specific training objectives.

4. Position within the diachronic experimental design

Hell-Char is paired with two temporally later evaluation sets that test transfer beyond the Hellenistic training period. PaLit-Char covers the 2nd–5th centuries CE, contains 384 images, is fully balanced, and is built from securely dated literary papyri in the PaLit dataset. It is structured as 4 specimens × 24 letters × 4 centuries; for the 5th century, where securely dated material was limited, additional paleographically dated material was used. Med-Char covers the 9th–14th centuries CE, contains 574 images, is built from Byzantine minuscule manuscript letters, and is balanced at roughly 24 images per letter per century where possible (Pavlopoulos et al., 23 Jun 2026).

Relative to Hell-Char, PaLit-Char represents the next major temporal stage and tests transfer across the Hellenistic-to-Roman transition. Med-Char introduces a much larger diachronic gap and tests whether learned representations survive the transition from Hellenistic cursive/majuscule forms to medieval minuscule forms.

This arrangement gives Hell-Char a specific epistemic role. It is not merely a standalone corpus; it is the historically earlier training anchor against which later robustness, temporal drift, and representation stability are measured.

5. Learning objectives and augmentations centered on Hell-Char

The paper proposes two methodological components trained on Hell-Char: a similarity-weighted supervised contrastive loss, termed DSCL, and a lacuna-driven augmentation scheme, termed LF (Pavlopoulos et al., 23 Jun 2026).

For anchor sample $i$, the supervised contrastive objective is

$$\mathcal{L}_i = - \frac{1}{|P(i)|} \sum_{p \in P(i)} \log \frac{\exp(\mathbf{e}_i \cdot \mathbf{e}_p / \tau)}{\sum_{a \neq i} w_{ia} \exp(\mathbf{e}_i \cdot \mathbf{e}_a / \tau)}$$

where $P(i)$ denotes positive samples with the same class as anchor $i$, $\mathbf{e}_i$ is the embedding of sample $i$, $\tau$ is the temperature, and $w_{ia}$ is a similarity-based weight for negative pair $(i,a)$. The core idea is that not all negative pairs are treated equally: the training procedure periodically recomputes an inter-class similarity matrix every 3 training epochs by passing the whole training set through the current model, computing class prototypes from normalized embeddings, taking cosine similarity between prototypes, clamping similarities to $[0,1]$, and setting diagonal entries to zero. An exponential moving average can stabilize updates. The stated effect is that very similar letters are not forced far apart, while dissimilar letters are pushed apart more strongly.

LF addresses manuscript damage. For each image, the procedure converts to grayscale, normalizes and resizes to $$\mathcal{L}_i = - \frac{1}{|P(i)|} \sum_{p \in P(i)} \log \frac{\exp(\mathbf{e}_i \cdot \mathbf{e}_p / \tau)}{\sum_{a \neq i} w_{ia} \exp(\mathbf{e}_i \cdot \mathbf{e}_a / \tau)}$$0, applies standard augmentations—rotation up to $$\mathcal{L}_i = - \frac{1}{|P(i)|} \sum_{p \in P(i)} \log \frac{\exp(\mathbf{e}_i \cdot \mathbf{e}_p / \tau)}{\sum_{a \neq i} w_{ia} \exp(\mathbf{e}_i \cdot \mathbf{e}_a / \tau)}$$1, translation, resizing, and color jittering—and then applies lacuna masking. The lacuna masks consist of 1 to 4 lacunae per image, each covering 2–15% of the image area; shapes are anisotropic ellipses, contours are distorted using erosion/dilation, masks are placed randomly, and masked pixels are replaced with background values. The paper contrasts this with random erasing and argues that LF better simulates flaking, humidity damage, worm holes, irregular surface loss, and ink absence due to physical degradation.

The main backbones evaluated on Hell-Char are a lightweight CNN, denoted fCNN, and ResNet18 variants trained from scratch or pretrained and fine-tuned. Preliminary experiments with ConvNeXt-V2 and ViT-S16 are also mentioned, but are not the main focus.

6. Recognition results, embedding structure, and paleographic interpretation

On Hell-Char recognition, the reported results are as follows (Pavlopoulos et al., 23 Jun 2026):

Model	Accuracy	F1
fCNN baseline	0.742	0.74
fCNN + random erasing	0.768	0.75
fCNN + LF	0.782	0.77
ResNet18 from scratch	0.788	0.74
ResNet18 pretrained + fine-tuned	0.801	0.79
ResNet18 pretrained + fine-tuned + SCL	0.818	0.81
fCNN + LF + DSCL	0.803	0.80
ResNet18 pretrained + fine-tuned + LF + DSCL	0.831	0.83

The best performer is ResNet18 pretrained + fine-tuned + LF + DSCL. The results also show that the lightweight CNN improves substantially with the proposed methods.

The learned embeddings are evaluated not only by recognition scores but also against raw pixels, Otsu thresholding + PCA, generic pretrained ResNet18 features, and pretrained but not task-adapted features. Clustering is assessed with k-means, spectral clustering, and agglomerative hierarchy, using NMI and ARI. The paper reports that ResNet18 + LF + DSCL gives the strongest clustering coherence by far; with k-means it achieves around NMI 0.768 and ARI 0.690, while Otsu + PCA performs much worse and pretrained ResNet18 alone is near-random.

These embeddings are then used for downstream paleographic analysis. The learned space supports clustering, identification of stylistic subgroups, and prototype image construction. For each $$\mathcal{L}_i = - \frac{1}{|P(i)|} \sum_{p \in P(i)} \log \frac{\exp(\mathbf{e}_i \cdot \mathbf{e}_p / \tau)}{\sum_{a \neq i} w_{ia} \exp(\mathbf{e}_i \cdot \mathbf{e}_a / \tau)}$$2 group, the prototype is a medoid, defined as the representative image closest to the group centroid in embedding space. Using spectral clustering and silhouette analysis, the paper identifies subtypes within letters; for Alpha, the model finds more than two clusters, including filled-in, circular, ligatured alpha, circular non-ligatured alpha, and angular alpha. The paper describes this partition as paleographically plausible.

Diachronic visualization is performed with t-SNE and one prototype per letter-century group. The resulting plots reveal gradual graphical change, transitional forms, temporal drift, letters that remain stable across centuries, and letters that split into distinct stylistic neighborhoods. The examples discussed include Iota, Lambda, and Chi as relatively stable; Gamma as shifting strongly from Hellenistic to medieval form; Beta as developing a minuscule, u-shaped form absent from Hell-Char; Zeta as showing “3-like” versus modern-looking forms; and Kappa and Nu as mixing older and minuscule forms.

Taken together, these results position Hell-Char as the historical anchor for a computational paleography pipeline in which domain-specific similarity modeling and domain-informed corruption modeling yield embeddings that are better for recognition, better for clustering, more faithful to paleographic structure, and useful for diachronic analysis across centuries.