Skip to content

Mismatch between valid scoring during training and real evaluation #2309

@vince62s

Description

@vince62s

We have a mismatch because:
The reference is made of the detokenized of tokenized target.... I know it should not be an issue at first glance.

The thing is that when you tokenize the target, you can get "<unk>" tokens and when you detokenize you lose the information.
As a result the scoring metric is in most cases better than it should be.

@l-k-11235 @anderleich
If one of you wants to fix this, welcome.

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