nah, it's better to train on all languages. 24 partitions? you are gravely underestimating these models and how they represent things in their latents... transfers easily