Change the Bisect Schedule to take a desired difference from baseline loss, rather than directly the desired loss
Change the plotting defaults (in case of ce_loss) to set y axis relative to baseline loss
Motivation and Context
Obviously better than having to guess a threshold
Previous defaults were always wrong
How Has This Been Tested?
Made some plots
Does this PR introduce a breaking change?
Yes. Existing BisectSchedule configs break. I updated the tests; otherwise we had none is main.
I think backwards compatibility is not worth it because score_target was a terrible argument and had to be adjusted every time we changed the data set.
Ablations with reference to baseline
Description
Motivation and Context
How Has This Been Tested?
Made some plots
Does this PR introduce a breaking change?
Yes. Existing BisectSchedule configs break. I updated the tests; otherwise we had none is main.
I think backwards compatibility is not worth it because score_target was a terrible argument and had to be adjusted every time we changed the data set.