Skip to content

AI feedback using logprobs #879

@tom-doerr

Description

@tom-doerr

Whenever I use a LM to evaluate something I ask a Yes/No question and do logprob_yes - logprob_no. This worked very well for me in the past. How can I access logprobs in my metric function?

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions