Skip to content

mode collapse issue #16

Description

@zhenzey

Thanks for the nice code! I was wondering whether there is a mode collapse issue witnessed in the generation. For instance, for the greeting case, the generated sample always starts with i's or i'm, but the original dataset seems more diverse.

i's that to see
i's nice are see you right
i'm let everyone you
i's stopped a you you!
i's glad hi see you right
i's glad to you
i's that to you you!
i's glad to you you right
i've chance to you you!
i's greetings a you

I have also tried my own dataset (6000+ data) and the mode collapse remains. Do you have any idea what might is going wrong? Or it is simply because of the size of the dataset?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions