Computer vision models are increasingly relevant and useful to Digital History. Next to the increasingly complex neural models, data and data selection are an integral part of this process. In this paper, we examine and extend the data collection practices from a major recent paper in the domain of architectural element classification. We collected an image-text data set for a selection of 56 Baroque landmarks to be analysed in like manner. This different architectural domain yielded insights into the transferability of the original model and data collection procedures. Notably, the architectural domain also has an impact on the availability of classes of architectural elements as well as the performance of the models classifying them.