sbert-spearman-not-pearson-metric

Summary

When evaluating how well a sentence-embedding model captures semantic similarity, the accepted metric is Spearman rank correlation, which checks whether the model agrees with human judges on the ordering of similarity scores rather than on exact numerical matches. This sets the standard for the system: a model is "good" if it ranks pairs consistently with humans, not if its raw scores align in magnitude, and Pearson correlation is explicitly rejected as the wrong tool for the job.