Silo AI and TurkuNLP are developing a family of multilingual open source LLMs with the aim of strengthening Europe's digital sovereignty and democratising access to LLMs. Developing baseline models that are in line with European values is crucial to this effort to ensure that they are based on data and information that accurately represents the different languages, citizens, organisations and cultural landscape of the European Union. This approach is not only in line with European values, but also enables sovereignty over how downstream applications and value are created. The success was attributed to the combination of the low-resource Finnish language with resource-rich languages. The team worked to determine the optimal frequency of data reuse for low-resource languages during training, and integrated translated text pairs from English and Finnish. This strategy, which relies on a cross-linguistic signal to improve the model's understanding of the relationships between the languages, proved crucial in achieving excellent performance in the low-resource languages without compromising performance in English.

Ask this episode

“What did the council decide on item 42?” — answered with the speaker and the moment.

Transcript

Related