Skip to content

How to reset entropy context -- avoiding entropy "drifting"? #135

@zdxdsw

Description

@zdxdsw

In $\S4.4$ of the paper, it says we reset the entropy context with new lines and use approximate monontonicity constraint as it suffers less from "entropy drift" from changes in context length.

I'm observing puzzling behaviors which I suspect could be due to how I treat "\n" during preprocessing. When I'm pondering the reason, I re-read the paper and found that entropy drifting might be related.
So I'm wondering about the implementation details of resetting the entropy context.

E.g. do you simply text.split("\n") before feeding the text to entropyLM -- this would incur lots of padding and de-padding? Or are there any other tricks?

Thank you in advance!!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions