It'd be nice to incorporate the '--truncate-source' flag to the 'multilingual_translation' task.
The '--truncate-source' flag already exists in the base 'translation' task.
Not sure why it was not incorporated into the 'multilingual_translation' task, but the change looks pretty straightforward, since the latter borrows the dataset creation method from the former:
from fairseq.tasks.translation import load_langpair_dataset
Sounds reasonable to me. Feel free to submit a PR.
Most helpful comment
Sounds reasonable to me. Feel free to submit a PR.