How does this work?

It seems like it, in most ways, it would be bad to train on 24 separate languages. That's just 24 partitions to the data. Seems really inefficient and better to simply train in the biggest (english) and translate.

I do think this will introduce some biases that correlate with the English language. It would be interesting to see more specifically what this means. But regardless, I don't think you can produce a competitive model with such a large subdivision of training data.