Problem About domain transfer with Augmented SBERT

[svjack]

I review the code located in train_sts_qqp_crossdomain.py The idea is to use cross_encoder's higher performance. So for unlabeled data, is it feasible to combine multi sample methods with cross_encoder's prediction and combine them with snorkel: https://github.com/snorkel-team/snorkel ? Do you have some advice for the design of label function about this labeling problem?

[nreimers]

I never used snorkel, so I cannot help there