PaLit-Char: Evaluation Dataset for Greek Letterforms

• PaLit-Char is a balanced evaluation dataset of 384 ancient Greek majuscule images from the 2nd–5th centuries CE, designed for diachronic testing.
• The dataset acts as a controlled stress test, contrasting late Hellenistic cursive training data with formal calligraphic majuscule bookhands to assess model generalization.
• Empirical results using models like ResNet18 show robust performance on PaLit-Char (accuracy ~0.84), highlighting its role in measuring enduring structural letter properties.

Searching arXiv for papers on PaLit-Char and related diachronic Greek letterform representation learning. PaLit-Char is a fully balanced evaluation dataset of ancient Greek handwritten character images drawn from majuscule literary papyri dated between the 2nd and 5th centuries CE. Introduced alongside Hell-Char and Med-Char in the study of diachronic representation learning for Greek letterforms, it is designed not as a training corpus but as a test set for assessing whether models trained on late Hellenistic material learn stable structural properties of letters rather than overfitting specific scribal habits or period-specific surface forms (Pavlopoulos et al., 23 Jun 2026).

1. Definition and position within the experimental design

PaLit-Char occupies the middle position in a three-part temporal setup: Hell-Char spans the 3rd–1st centuries BCE, PaLit-Char spans the 2nd–5th centuries CE, and Med-Char spans the 9th–14th centuries CE. Within that design, PaLit-Char is used exclusively as an out-of-training-distribution evaluation set for diachronic generalization from late Hellenistic cursive material to Late Antique bookhand letterforms (Pavlopoulos et al., 23 Jun 2026).

The dataset is explicitly motivated as a “chronologically close” but stylistically different target domain. Hell-Char covers late Hellenistic cursive documentary hands, whereas PaLit-Char consists of calligraphic majuscule literary bookhands. This makes PaLit-Char a controlled stress test: the chronological jump is modest, but the script style changes from documentary cursive to formal calligraphic majuscule. In the paper’s framing, this intermediate position also provides a contrast with Med-Char, where the chronological and graphical jump is much larger because the later material is Byzantine minuscule.

A recurrent misconception is to treat PaLit-Char as a general-purpose historical Greek handwriting corpus. It is more specific than that. Its role is evaluative and diachronic: it measures whether a representation learned from earlier material remains robust when faced with later but still recognizably related letterforms.

2. Construction, balance, and annotation

PaLit-Char is derived from PaLit, a dataset of securely dated literary papyri introduced in earlier work by some of the same authors. For the 5th century, where securely dated literary papyri are scarce, 48 images were taken from one additional paleographically dated manuscript. The source base is therefore literary papyrological rather than documentary, and its core script type is calligraphic majuscule bookhand (Pavlopoulos et al., 23 Jun 2026).

Aspect	Value	Note
Total images	384	Fully balanced
Letter classes	24	Standard Greek letters
Centuries	4	2nd–5th centuries CE
Specimens per letter per century	4	Balanced by design
Samples per letter	16	$4$ centuries $\times 4$ specimens
Samples per century	96	$24$ letters $\times 4$ specimens

The paper describes the dataset as “exactly balanced” across letters and centuries: 384 images, arranged as $4$ specimens $\times 24$ letters $\times 4$ centuries. There are no numerals or “Unknown” symbols; the dataset is strictly letter-level. Each sample carries a letter label, and each sample also has an associated century label used for diachronic analysis, although the recognition models predict letter identity rather than century.

All character images are brought into a common preprocessing pipeline. Each image is converted to grayscale, normalized, and resized to $64 \times 64$ pixels. During training phases that involve the same modeling framework, the augmentations include rotation up to $10^\circ$, translation, resizing, color jittering, and lacunae-inspired masking. The paper does not restate a PaLit-specific segmentation pipeline in detail, but it treats PaLit-Char as character-level material already compatible with this preprocessing regime.

The dataset is small, curated, and label-focused rather than large-scale and noisy. No crowdsourcing is mentioned. The use of securely dated sources for most centuries, together with paleographically dated supplementation for the 5th century, places it closer to an expert gold-standard evaluation set than to a large observational corpus.

3. Paleographic profile and diachronic significance

PaLit-Char captures a transitional but still strongly majuscule phase of Greek handwriting. The broader diachronic background described in the paper is that epigraphic letterforms begin to be modified with increasing cursivity in the Hellenistic period, while calligraphic writing styles preserve capital-like forms in literary contexts. PaLit-Char belongs to that calligraphic majuscule or bookhand stream, whereas Hell-Char belongs to late Hellenistic cursive documentary writing, and Med-Char belongs to Byzantine minuscule (Pavlopoulos et al., 23 Jun 2026).

This distinction matters because PaLit-Char is not merely “later Hell-Char.” It is a different scribal regime. Compared with the late Hellenistic documentary hands of Hell-Char, PaLit-Char contains more regular, standardized, and legible forms. At the same time, it remains close enough to earlier capital-derived letter structures that it can still function as a test of diachronic continuity rather than wholesale script discontinuity.

The paper situates PaLit-Char as the last major majuscule stage before the minuscule transformation represented by Med-Char. Some letters remain comparatively stable across all three datasets, including Chi, Iota, and Lambda. Others diverge sharply only in the later medieval stage. Examples emphasized in the paper include Gamma, Beta, Zeta, Omega, Kappa, Mu, Nu, and Upsilon. In this sense, PaLit-Char is paleographically important because it preserves a high degree of continuity with capital forms while still recording systematic stylization and variation.

A plausible implication is that PaLit-Char isolates script-style shift more cleanly than Med-Char does. Hell-Char to PaLit-Char tests whether a model can generalize across cursive versus calligraphic execution within a still-majuscule regime; Hell-Char to Med-Char additionally tests recognition across a much deeper structural reorganization of letterforms.

4. Role in representation learning and model design

The main modeling setup trains on Hell-Char and evaluates on PaLit-Char. The backbones are a lightweight CNN and a pretrained ResNet18, with the strongest reported results coming from ResNet18 fine-tuned with lacuna-driven augmentation and a similarity-weighted supervised contrastive loss termed DSCL. In the principal experiments, PaLit-Char is not used for training; it is used to assess out-of-temporal-distribution generalization (Pavlopoulos et al., 23 Jun 2026).

The supervised contrastive component is defined per sample by

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

Here $\times 4$0 is the embedding of sample $\times 4$1, $\times 4$2 is the set of same-class positives, and the denominator weights negatives according to dynamically estimated inter-class similarities. The paper’s central idea is that visually similar letters should not be forced apart as aggressively as unrelated letters. Class prototypes are recomputed from embeddings every three training epochs, cosine similarities between prototypes produce a $\times 4$3 similarity matrix, and that matrix governs the contrastive weighting.

The augmentation scheme most relevant to PaLit-Char is lacuna-driven fragmentation. Rather than using only generic erasures, the model is trained with irregular masked regions intended to approximate lacunae observed in damaged manuscripts. The number of lacunae is sampled between $\times 4$4 and $\times 4$5, and each covers between $\times 4$6 and $\times 4$7 of image area. Shapes begin from anisotropic ellipses and are altered by random morphological operations to resemble organic damage patterns such as flaking or wormholes.

PaLit-Char is therefore not only a test of raw recognition. It is also a test of whether these inductive biases—similarity-aware contrastive structure and domain-informed corruption—yield embeddings that remain coherent under moderate diachronic shift.

5. Empirical performance and representation behavior

The strongest reported PaLit-Char result comes from the ImageNet-pretrained ResNet18 fine-tuned on Hell-Char with lacuna-driven augmentation and DSCL. On that setup, PaLit-Char achieves $\times 4$8 accuracy and $\times 4$9 F1, essentially matching or slightly exceeding the corresponding Hell-Char test performance. For comparison, the same model drops sharply on Med-Char, where the reported values are $24$0 accuracy and $24$1 F1 (Pavlopoulos et al., 23 Jun 2026).

Test set	Period	Accuracy	F1-score
Hell-Char	3rd–1st c. BCE	0.83	0.82
PaLit-Char	2nd–5th c. CE	0.84	0.84
Med-Char	9th–14th c. CE	0.45	0.42

The paper notes that some letter-specific F1 values decrease on PaLit-Char, including Phi, Pi, and Psi, while others improve, including Alpha and Zeta. The authors attribute the overall strength of PaLit-Char performance to the calligraphic nature of the material: regular, standardized, and legible forms appear to offset the temporal and stylistic shift away from the late Hellenistic training domain.

The embedding analysis in the paper is most explicit for Hell-Char and Med-Char, but the PaLit-Char results are informative precisely because they are obtained without direct training on that dataset. This suggests that the learned representation space preserves enough diachronic continuity that PaLit-Char instances fall into the correct letter regions despite being later and stylistically different. The paper does not provide a dedicated mixed t-SNE plot for Hell-Char and PaLit-Char together, nor a full PaLit-Char confusion matrix, but the classification results strongly indicate effective transfer.

The authors also report an intermediate fine-tuning experiment: fine-tuning on PaLit-Char before evaluating on Med-Char raises Med-Char accuracy only from $24$2 to $24$3. This is a modest gain. It confirms that PaLit-Char can function as a bridge domain, but it also shows that the later shift into minuscule script is not resolved by a single intermediate stage of Late Antique majuscule adaptation.

6. Limitations, misconceptions, and research uses

PaLit-Char is small. It contains only 384 images, with 4 examples per letter per century. It is therefore well suited to controlled evaluation, but not to large-scale standalone training. The paper also emphasizes that the corpus is script-selective: it contains literary majuscule papyri and predominantly formal bookhands rather than the full spectrum of Late Antique handwriting (Pavlopoulos et al., 23 Jun 2026).

A second limitation is chronological and source bias. The 5th-century component is supplemented by one additional paleographically dated manuscript because securely dated material is scarce. This may overrepresent a narrow scribal or regional tradition. A third limitation is that the dataset is organized by letter and century rather than by scribe, so it is not directly optimized for authorship or scribal-attribution studies.

Several misconceptions can therefore be dispelled. PaLit-Char is not a comprehensive survey of Greek handwriting in the 2nd–5th centuries CE. It is not intended to model the full transition to minuscule. It is not a large benchmark for training foundation models. Its design goal is much narrower and more precise: balanced, expert-curated, diachronic evaluation at the character level.

Within that scope, the dataset has several research uses. It functions as a fixed benchmark for testing diachronic robustness of character embeddings; as an auxiliary character-level resource for OCR and HTR on Roman and Late Antique Greek literary manuscripts; and as a paleographic comparison point between Hellenistic cursive and medieval minuscule. The paper also presents prototype-based visualization and clustering methods for letter evolution, and PaLit-Char naturally serves as the Late Antique middle stage in such trajectories. Code and data are reported as available through the project repository linked in the paper (Pavlopoulos et al., 23 Jun 2026).

In the larger context of diachronic representation learning, PaLit-Char is valuable because it isolates a specific kind of generalization problem: moderate temporal distance, clear stylistic reformatting, and limited but high-quality supervision. That design makes it a precise instrument for measuring whether learned representations capture enduring graphemic structure rather than merely period-bound appearance.