I notice this one is slow to get off the ground, and after starting work on a starter notebook I can see why! Unlike some other, more common NLP tasks translation is still pretty complicated, and most tutorials rely on pre-trained LMs for specific languages, tons and tons of code that you need to copy/paste and/or obscure training scripts that are super hard to understand. I have yet to complete a proper notebook or submission but in the meantime, I thought I'd share some useful resources:
1) A great multi-part series by Stephen Oni (first part https://heartbeat.fritz.ai/exploring-language-models-for-neural-machine-translation-part-one-from-rnn-to-transformers-3e53b7d8a01f, code: https://github.com/steveoni/English_Yoruba_Transformer). This is very comprehensive, starting with basic concepts and explaining some of the latest techniques used in transformer models.
2) BLURR: Fastai + transformers for Seq2Seq tasks. Let's you use fastai goodness for dataloaders and training. ps://ohmeow.github.io/blurr/modeling-seq2seq-translation/
3) SimpleTransformers recently added translation. Good blog post with an example here: https://towardsdatascience.com/how-to-train-an-mt5-model-for-translation-with-simple-transformers-30ba5fa66c5f
I'm hoping to have a nice starter notebook ready to share soon but with all that is going on, there is no guarantee of anything! May the above be useful to you in forging your own path, and please share what you can to help us all move this area forward :)