Transfer learning with mbart

Support
1

Hello,
I am using the mbart model as a pretrained model for machine translation with Target language not existing in the list of dicts.
Is it possible to use train_from (with opennmt_py) and finetune it on different languages?
I keep getting this error:
fields = checkpoint[‘vocab’]
KeyError: ‘vocab’

Any advice?

2

I believe the answer is in this topic:

https://forum.opennmt.net/t/fine-tuning-mbart/4700

2 Likes
3

Thanks Samuel!

I think that the topic you are referring to is fine-tuning for a language pair that is already available in MBART. Is there any additional steps that needs to be done to fine-tune for a new language pair or to add a new Target language as @cha mentioned?