Iconclass Annotations and LLM-Generated Descriptions for Selected Wenzelsbibel Miniatures

This dataset accompanies the article Transferring AI-Based Iconclass Classification Across Image Traditions: A RAG Pipeline for the Wenzelsbibel (https://doi.org/10.3390/histories6010017) by Drew B. Thomas and Julia Hintersteiner, published in Histories 6:1. It contains scene-level data for selected miniatures from the Wenzelsbibel, including IIIF links to images, large language model–generated textual descriptions, expert-assigned Iconclass ground-truth annotations, and Iconclass codes predicted by a multimodal retrieval-augmented generation (RAG) pipeline. All descriptions were generated using GPT-4.1 with a full-page image input and an “Old Testament illustration” prompt. For page-level description generation, only manuscript pages containing a single miniature were used to ensure unambiguous correspondence between images and annotations. Predicted Iconclass codes correspond to the best-performing configuration of the pipeline (Phase 2.2b), which integrates hierarchical reasoning and a refined rescue strategy. The dataset is released to support transparency, inspection, and reuse of the results reported in the article and does not include intermediate experimental configurations.