On Serval QA, any text that has already been entered by the user is not being pre-translated. This is a regression of #370.
Compare this pre-translation on Serval Prod:
To Serval QA:
Both builds had the same source and target, and were both set to train on MRK and pre-translate JHN. The translations were generated with only 20 max_steps.
Pre-translation of already written text is an important feature of Serval (and was originally requested by users), so I would classify this as a blocker for QA going to production.
On Serval QA, any text that has already been entered by the user is not being pre-translated. This is a regression of #370.
Compare this pre-translation on Serval Prod:
To Serval QA:
Both builds had the same source and target, and were both set to train on MRK and pre-translate JHN. The translations were generated with only 20
max_steps
.Pre-translation of already written text is an important feature of Serval (and was originally requested by users), so I would classify this as a blocker for QA going to production.