setfit for ranking problems?

I'd like to finetune setfit for ranking. Where, for each instance, there are possible text anwers in a ranked order.

I'd then like to finetune to maintain the ranking. Would this be possible?

[fast-soft-sort](git@github.com:google-research/fast-soft-sort.git) can be used for differentiable ranking.

The idea might be that I am not fine-tuning the sentence-transformer of the input (or I can, optionally), but the output similarity transformer would be fine-tuned to optimize the rank score.

No files, tests, or entry points are named. First determine how SetFit's output similarity transformer is trained and whether fast-soft-sort can provide a differentiable ranking objective; done would require a maintainer-approved design for the training scope and ranking loss.

Written by the indexing model from the issue text.