Skip to content

Benchmark evaluation fails for Yes/No questions due to label mismatch #3

@cjfcsjt

Description

@cjfcsjt

The current benchmark evaluation logic cannot correctly handle Yes/No questions when there is a mismatch between prediction format and ground truth labels.

Model prediction (pred): "yes" / "no"
Ground truth (GT): "A" / "B"
Because of this inconsistency, correct answers are being marked as incorrect during evaluation.

Steps to Reproduce
Run the evaluation on Ego_Centric_Abosolute_Distance
Observe that correct predictions are not counted as correct.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions