This paper introduces a new method for classifying medieval Hebrew Manuscripts’ script types(writing styles) using a semi-supervised deep-learning approach. The approach is based on a Siamese network that provides learned features of unlabeled datasets. The learned features are used to construct clusters representing each type of script. We conducted a detailed ablation study using state-of-the-art feature extractors with various configurations to pick the best model for extracting learned features. At the end of our experiments, we provide an analysis of our method that provides results comparable to the state-of-the-art approach.