Specifying ones own samples or sentence pairs

I have been looking into SetFit for a while and its a very nice tool. Currently, it seems that there are a lot of options for training and using the classification head, but there is no way to train the 'body' on a specified set of label pairs. Would it be possible or appreciated if such capacity was added?

One of the key reasons is that currently all samples/classes are treated the same when doing contrastive learning (correct me if wrong). It would be nice to specify the pairs myself as to enhance the contrastive capacity (hard negatives / positives).

If there are already options for this let me know.