Recent work to use MinT to provide initial translations for the Translate extension (T338131), including its use in Translatewiki.net (T340544) can be complemented by exporting the final translations into a dataset and integrating it into the Opus project.
In this way, the corpus of multilingual text can be expanded with the data form localization strings and translatable pages. Resulting in more data to train the next version of the models.
As an initial step we may want to generate some samples that can be helpful to coordinate with the Opus team and make sure the format provided is a useful one.
In a similar effort, published translations from Content translation are already integrated into Opus.