The SAE was originally trained by Gytis Daujotas. You can read more about his work here. The model has been adapted to work with Prisma to facilitate further research and application within the Prisma ecosystem.
Note: In the adaptation process, the weights were transformed to accommodate Prisma’s handling of the SAE forward pass. As a result, the model’s output may exhibit slight numerical differences compared to the original. These discrepancies should not affect the overall performance or interpretability of the model.