30 manuscript-catalogue card images from the Duke Rubenstein collection
(biglam/rubenstein-manuscript-catalog),
used as a held-out generalization probe: a collection and schema the
index-card-extractor-4b model was never trained on. Tests whether the model still emits
valid, schema-conforming JSON for an unseen schema (it does — better than the base model).
No ground truth — this measures schema-following (valid-JSON rate… See the full description on the dataset page:
https://huggingface.co/datasets/davanstrien/rubenstein-probe-input.
