Should we use large language models, based on major languages, or should we develop our own? AI Sweden believes that developing our own models is more sustainable, and thus, they created the large-scale generative language model GPT-SW3. Trained on Swedish, Danish, Norwegian, Icelandic, and English data, the model performs well in Nordic languages. Training with several languages leverages cross-lingual transfer learning. As the model has not been trained on Faroese data, it was a positive surprise for us to find the model performing quite well in understanding Faroese. These initial findings have led to researching the GPT-SW3 capabilities for Faroese further and investigating how we can leverage the model for Faroese purposes.
In connection to this, Kringvarp Føroya interviewed Iben Nyholm Debess about GPT-SW3. The segment can be heard here.
