Language models for Swedish authorities
Reference number | |
Coordinator | Lindholmen Science Park AB - AI Sweden |
Funding from Vinnova | SEK 6 635 029 |
Project duration | November 2019 - October 2022 |
Status | Completed |
Venture | AI - Leading and innovation |
Call | From AI-research to innovation |
Important results from the project
The project´s main objective has been to promote the development and use of language models in the public sector in Sweden. To achieve this objective, the project has developed and contributed prerequisites and resources, including knowledge sharing, data, models, code, use cases and applications. The objective has been met by a large margin. All authorities that participated in the project have developed and evaluated solutions based on Swedish language models, and the same applies to other actors in the public sector who were part of the project´s reference group.
Expected long term effects
The project results include the language models that are now available for Swedish (most of which originate in the project), the code and the academic publications that have been published, and the applications that have been developed by both project partners and actors within the reference group. Furthermore, the project has contributed to Swedish language technology being at the forefront of developments in language models. The project has been the driving factor in the Swedish public sector becoming a leader when it comes to applying Swedish language models in concrete use cases.
Approach and implementation
The work has been carried out in five work packages, where WP1 focused on data and evaluation, WP2 focused on algorithms and model types, WP3 has dealt with implementation and operation, WP4 has focused on applications, and WP5 has dealt with coordination and dissemination. The project has acted as a common ground on which various activities have been carried out by the participating parties, such as model training, experimentation, application and evaluation. Collaborations have consisted partly of sharing resources (code, data and models) and partly of joint development projects.