How can I finetune an embedding model with a multi label dataset for similarity comprison?

I have a multi-label dataset with two columns, sentence and label. I want to train an embedding model for similarity comparison .
Currently, I am training with the following steps:

  • loading the origin dataset.
  • pairing the sentences, assigning similarity 1 to pairs with the same label and 0 to others.
  • constructing a new dataset with three columns sentence1, sentence2, and score
  • finally train the model with the new dataset

I found the pairing process time consuming. And it would consume a lot of disk space if I store the new dataset as there are many duplicate sentences. Is there any solution to avoid the pairing process while consuming low disk space?