Hello,
I was wondering if its possible to finetune MBart on summarization (src_lang=tgt_lang) in one language and zero-shot transfer to another language ? In the same fashion as if you would finetune MT5 in one language on a downstream task it also learns to do it for other languages. Would the model recognize that any monolingual pair should be summarized or it would only work on the language token id of the fine tuning ?
Thanks !