Data without negative examples

Hi

I was studying active learning integration using setfit with argilla

Turns out I was using a small dataset, only 5 examples in each round, i.e. active learning trained the setfit with 5,10,15,20 examples according to the manual labeling pace.

In the case of a multi-label problem, in some cases there were no negative data in the small set, that is, they all had the same label in common.

This way it was not possible to generate the positive negative pairs, so data was not generated for training and thus a raise occurred.

My question is knowing that setifit needs data where the label doesn't exist to generate the pairs, what should I do? Ensuring there is always a negative example in the set, as this is user-labeled data this could be a problem. Or increase the set of data that are sent at a time, instead of 5 put 50 for example. However, there is always a slim chance of it happening again.

`src/setfit/logging.py` in sentence_pairs_generation_multilabel this is where sentence generation takes place.