I think the reasons the proposed German model wasn’t very good are:
- The current German model is trained with OpenNMT-tf, which is more production oriented, while the proposed one was trained with OpenNMT-py, which is more research oriented.
- I used the CCMatrix dataset from Opus in the proposed model. I think this dataset is large and low quality and I now usually avoid it.
I’m planning to keep the current model for now because it’s high quality and trained with OpenNMT-tf while I’ve since moved my training infrastructure to OpenNMT-py. OpenNMT-tf models are also around 30% larger but that’s worth it in this instance because German-English is a very popular language pair. If good German CTranslate2 models from third parties become available I may consider switching to them too.