If the language has been included during pretraining, I that shouldn't be a problem (via swapping the dataset). Otherwise, I am not sure:

→ View original post on X — @rasbt