Cohere's RCP-nDCG Metric Raises Human Agreement to 97%

Cohere's Nils Reimers shows traditional nDCG gains only 53% human agreement—near coin-flip—and introduces RCP-nDCG@10, used to train Embed 5, lifting human agreement to 97% in real-world retrieval.

2026-10-03 ~ 2026-10-03 · 3 related posts

• Cohere Embed 5 Is First Model Family Benchmarked With New RCP-nDCG@10 Metric — cohere · 2026-10-03
• Traditional nDCG agrees with humans only 53% of the time, RCP-nDCG hits 97% — Nils_Reimers · 2026-10-03
• Cohere says its RCP-nDCG metric fixes embedding evals: 97% human agreement vs 53% for nDCG — Nils_Reimers · 2026-10-03