At the moment validation bleu barely gets above zero in the tests, so they don't really prove much about our code.

we could use a larger model like sshleifer/student_marian_6_3, and more data, and train for 10 minutes . This would allows us to test whether changing default parameters/batch techniques obviously degrades performance.

The github actions CI reuses it's own disk, so this will only run there and hopefully not have super slow downloads.