How to fine-tune a subset of the vocabulary?

[danyaljj]

Hi there —

For a fixed model (say, GPT-2) how can I fine-tune only a subset of the word embedding? (say, a few newly added tokens added to the vocabulary)

Thanks!