When invoking the --prefix arg from the CLI, this was prepended to the FIM prompt. This works fine for models like incoder, which don't explicitly define FIM mode with a token at the start of the prompt, but with the bigcode models, it leads to weird behavior. I've adjusted such that the CLI prefix arg is prepended before the FIM prefix code but after the <fim_prefix> special token. This results in comparable accuracy on DS-1000 with and without --prefix being specified at the CLI.
When invoking the
--prefix
arg from the CLI, this was prepended to the FIM prompt. This works fine for models likeincoder
, which don't explicitly define FIM mode with a token at the start of the prompt, but with thebigcode
models, it leads to weird behavior. I've adjusted such that the CLI prefix arg is prepended before the FIM prefix code but after the<fim_prefix>
special token. This results in comparable accuracy on DS-1000 with and without--prefix
being specified at the CLI.