Spaces:
Running
Add Hy-MT2 series and TranslateGemma
Hello, how about add Hy-MT2 series and TranslateGemma? They are also very strong models.
Hi @hugeface99 ! Thanks for the suggestion!
From what we have observed, in terms of the language coverage, TranslateGemma is not much better than the original Gemma3: the MT-specific post-training has improved the fluency and robustness of translations in the languages that the models already knew, but didn't really increase its language coverage. And because BOUQuET mostly consists of sentences that are easy (for the models that do know the languages well), the improvements of TranslateGemma on the already-supported languages won't be very well visible on this benchmark. So I won't expect this model to shine — but could still include it for the sake of completeness.
As for Hy-MT2 — this series is interesting and we haven't touched them yet, but I will try including them into the benchmark in the foreseeable future!