Comment by drawfloat
3 months ago
If I ask a model to transcribe something exactly and it outputs an interpretation, that is an error and not a success.
3 months ago
If I ask a model to transcribe something exactly and it outputs an interpretation, that is an error and not a success.
Author already mentions that a correction is still an error in the context of this task.
And then refers to it as almost perfect. Being unable to follow a basic command like that means it is “nearly usable” rather than “nearly perfect”.