Swedish Lion — Riksarkivet's AI Model Making a Million Handwritten Historical Documents Searchable

Sweden's National Archives (Riksarkivet) operates its own AI research unit, AIRA, which built 'Swedish Lion', a handwritten text recognition (HTR) model trained specifically to read Swedish handwriting from roughly 1600 to 1900. The training data came from 'Transkriberingsnod Sverige', a three-year, publicly funded collaboration in which volunteer genealogy researchers manually transcribed historical documents to build the ground-truth dataset the model needed.

The project aims to make large historical archival collections full-text searchable, starting with two major collections: the 17th-century Witchcraft Commission archive and the 18th-century Svea Court of Appeal protocols.

Riksarkivet has published the machine-interpreted texts as open data, downloadable via API or as XML files, and previously open-sourced its underlying HTR/OCR tooling, including the 'HTRflow' framework, for reuse by other archives and researchers.

The model has made roughly 1.2 million pages full-text searchable in the National Archives Database, from an initial batch of over one million handwritten documents. Riksarkivet has said millions more archival pages will be processed and published progressively through 2025.

Because the project's own materials do not publish a formal, independently verified accuracy or error-rate figure, the practice is evidenced primarily by its delivered, publicly checkable output rather than by a benchmarked accuracy score.