Skip to content

IMDB Experiment Training Details #2

@MINGXUAN99

Description

@MINGXUAN99

Hi, thank you for this insightful and impactful work.

Could you please provide more details about the Small Scale experiment(IMDB)? In particular, I would appreciate clarification on how the positiveness and conciseness scores reported in Figures 2–4 are computed.

Additionally, I am a bit confused about the conciseness results. In Figure 4, the average conciseness score decrease after alignment with MaxMin-RLHF, whereas in Figure 5 it appears to increase. Are these conciseness scores computed in the same way, or different evaluation metrics are used?

Thank you so much for reading my questions!

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