Multilingual
#1
by
Gabriel
- opened
Hi! I tried to train FLAN-t5 for Swedish xsum (https://huggingface.co/Gabriel/flan-t5-base-xsum-swe).
However, it got poor rouge score for FLAN compared to when I trained a Swe Bart model (https://huggingface.co/Gabriel/bart-base-cnn-xsum-swe).
Do you know why this is the case?
Not sure why BART is performing better, but keep in mind that both models are not pretrained on the Swedish language.
I think it would be worth it to try a mT5 model, e.g. https://huggingface.co/google/mt5-base which was pretrained on some Swedish language as well.