Skip to content

Question regarding WISE implementation on Multimodality data #648

@boilerchun

Description

@boilerchun

Thanks for the work!

I have a question regarding the implementation of WISE on multimodality dataset.

If I understand correctly, the current implementation of WISE on multimodality dataset seems directly use the locality data (which used for testing/evaluation) during the training. Wouldn't this result in leaking the test information?

Also, as the comment points out, the loc_prompt is not equivalent to locality prompt. However, the implementation seems they are equivalent.

# TODO: locality_prompt(Unrelated Knowledge) is not equal to loc_prompt(Subject phrase)!

Metadata

Metadata

Assignees

No one assigned

    Labels

    questionFurther information is requested

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions