Skip to content

Latest commit

 

History

History
10 lines (7 loc) · 709 Bytes

README.md

File metadata and controls

10 lines (7 loc) · 709 Bytes

Do we need to freeze embeddings when fine-tuning our LM?

An experiment to check the impact of freezing embedding matrix on the validation metric depending on the number of non-overlapping tokens between train and validation set.

The experiment was run in the colab notebook, you can check the tensorboard logs there. There is more detailed explanation in the article. The main results can be viewed on the following plot:

main result plot