Leibniz Legible

Leibniz Legible is an access layer for the roughly 236,000 page images of the Leibniz Nachlass that the Gottfried Wilhelm Leibniz Bibliothek in Hannover has digitized and placed online. Every page is machine-read by a handwritten-text recognition model trained for Leibniz's hand, and every line carries its own confidence score and its provenance: which model read it, in which run, on which image. The result is searchable with tolerance for early-modern spelling and browsable in an open IIIF viewer that loads the library's images directly and shows the machine reading beside them, with honest labels on every line.

The project's register is deliberately modest. Its output is machine text with about an eight percent character error rate on Latin and French, higher on German; it is not an edition, and it is explicitly subordinate to the Akademie-Ausgabe, which has printed about a quarter of the Nachlass in a century of work. The point is to make the other three quarters legible and findable while the edition continues. Ground truth for improving the model is minted by aligning the reading text of copyright-expired edition volumes onto the machine transcriptions, at scale and with the alignment confidence kept on every line.

The code is open (Apache-2.0), the reports are published on Zenodo (CC BY 4.0) and mirrored in the Research section here, and the datasets and model follow with their own DOIs as the phases complete. The project is built on the PHILIUMM project's open model and ground truth (ERC, Paris), whose numbers it reproduced before building on them.