Which method you used to aggregate the pairwise scores?

The paper mentions "At inference time, we aggregate the pairwise scores pi,j so that each document receives a sin- gle score si. We investigate five different aggregation methods (SUM, BINARY, MIN, MAX, and SAMPLE)"

I didn't see this part of code.

Thanks for asking and sorry about not adding a comment mentioning it.

We use the SUM method:

duobert/run_duobert_msmarco.py

Line 443 in 24c7a16

pred_docs = scores.sum(1).argsort()[::-1]

Thanks!