The RSE-BERT-large-STS is trained with 2 relations including:
1) entailment
2) duplicate_question

The BERT-large-uncased model is used as initialization. 

It can be used ideally for STS datasets.