We recently discovered that we're only training on and aligning the content in the pretranslations in nmt_engine_build_job.py. We should be training on the training data and aligning the pretranslations. This will almost certainly greatly improve the quality of these alignments. Possibly for NLLB-out-of-the-box situations, we may want to train on the pretranslations since there's no training data, but in general since these are not guaranteed to be accurate, we should not be training on them.
We recently discovered that we're only training on and aligning the content in the pretranslations in
nmt_engine_build_job.py. We should be training on the training data and aligning the pretranslations. This will almost certainly greatly improve the quality of these alignments. Possibly for NLLB-out-of-the-box situations, we may want to train on the pretranslations since there's no training data, but in general since these are not guaranteed to be accurate, we should not be training on them.