Generative Language Model for the Judiciary (GSJ)

The project “Generatives Sprachmodell der Justiz” (GSJ) aims to develop a generative language model. It was initiated in early 2023 with the goal of training a language model on judicial data that has not previously been publicly available.

We aim to develop a model that can be operated locally, on the computing infrastructure of the German state justice administrations. Court files contain highly sensitive personal data that cannot be handed to external inference providers, which rules out the commercial API models that dominate current legal AI practice. The choice of open-weights models, the pseudonymization pipeline, the training setup etc. follows from that constraint.

The work divides into three pillars: making judicial data usable for model development in the first place, building an agentic-driven application that lets a model work with electronic case files (eAkte) the way a judge does, and training a model for the legal tasks that this work involves. Annotation of the released case files is carried out under the direction of our project partners at the University of Cologne.