why do we need + "<|endoftext|>" when processing query and doccument

#3
by Mengyao00 - opened

what if removing it? will it cause error or accuracy drop?

I think there will be accuracy drop here, as the model trained to use the EOS token as representation.

If you retrain/finetune it without adding the eos, (using avg pooling or simple last token) it should be fine.

Sign up or log in to comment