norkart/mt5-large-nn

This is a pruned version of the google/mt5-large model. Here, the input and output embeddings are pruned to support a greatly reduced vocabulary. The chosen vocabulary has 30K norwegian, english and special tokens, ~12% of the old size. This reduces the model size by roughly 37%. The model is still OK on similar languages, like German and Danish, but very different languages like arabic are not a good fit anymore. This model is intended as a starting point for finetuning mt5 for norwegian applications.

In addition it has been trained on Nynorsk using masked language modeling, in accordance with pretraining objective in the T5 paper.