Wondering the performance imporvements on larger language models.

[DreamInvoker]

Thanks for your great and effective work! Have you tried your method on 7B or 13B LLMs like LLaMa or Bloom? Thank you again.

[MattYoon]

Hi, @DreamInvoker. Thank you for your interest in our work!
Unfortunately, we have not. Preliminary experiments showed similar performance gains on GPT-Neo models of similar sizes. But nevertheless we did not test GAP on larger models due to computational limits.