Did you know if you carefully combine all the training best practices, you can get a SoTA retrieval model with less data and compute? @bclavie
, does, so he did, so here's the new SoTA for Japanese retrieval:

→ View original post on X — @jeremyphoward