Also enable use of Tatoeba-Challenge NMT models?

[jorgtied]

Could we even use the Tatoeba Challenge NMT models from https://github.com/Helsinki-NLP/Tatoeba-Challenge/tree/master/models? Does it work to use iso639-3 language codes? Can we even use models that cover language groups? (see https://github.com/Helsinki-NLP/Tatoeba-Challenge/blob/master/results/tatoeba-results-langgroup.md)

[xulihang]

I tried https://huggingface.co/Helsinki-NLP/opus-mt-en-zh and it works. But I have to change the language code to iso-639 (e.g. eng-> en) for Trados to use.

[TommiNieminen]

Currently only bilingual models are properly supported (even the multilingual OPUS models probably won't work correctly). There's already some adhoc conversion happening from ISO-639-2 (three-letter) to ISO-639-1 (two-letter), since Trados uses 639-2 and memoQ 639-1, but should make a proper facility for conversion between the different code schemes.

Related to the Tatoeba-Challenge models, I should add a possibility of defining multiple model repositories in the UI (right now it's just one repository, and the settings can only be changed in the config files) so that Tatoeba-Challenge repository can be used. That would also be useful for local model repositories for companies that customize models etc.

[TommiNieminen]

The new release of OPUS-CAT now support Tatoeba-Challenge models (including multilingual models), so I'm closing this.