Traditional nDCG only credits results already in a benchmark's answer key. RCP-nDCG@10 grades every result on its actual relevance, and its results are backed by human reviewers and MTEB.

→ View original post on X — @cohere