Use this topic to ask your questions to Lewis Tunstall during his talk: Simple Training with the Transformers Trainer
Is it possible to share the link to the notebook Lewis is working on here?
You suggest changing bert-base for MiniLM as baseline pre-trained model. Is there an analogous recommendation for multilingual models? Which checkpoint should be the go-to?
I added it on the first post
During the technical glitch Lewis explained the weighted loss
Can he show it again?
Is there a page in the docs / chapter in the course / an external resource with an overview of different loss functions and their advantages/disadvantages?
How can you use the dataset/trainer/pipeline approach with your own data?
Does the way the initialized model was made (as in made using tensorflow or pytorch) affect its performance or cause any issue?
This is answered at 25:00 in the main stream.
Like Sylvain and Lewis mentioned, this would be a great resource to have. I’d love to collaborate on this
This is shown at 27:05 in the main stream.
This is answered at 28:37 in the main stream.
This is answered at 29:37 on the main stream.