How the 7 negative samples are selected? Top7 below 0.95*postive_doc_score?
#20
by
YWang17
- opened
Hi, thank you for your the very solid work. It is mentioned that you used the positive-aware negative samples tech in Nv-Retrieval-v1. In that paper, it only selected the top4 docs below 0.95xPostive_doc_score. How is it on Nv-embed-v2? Still top7 after 0.95xpostive_doc_score as negs?