mteb-pair-classification-distance-metrics

Summary

Pair classification in MTEB doesn't lock in on a single similarity measure; it tests cosine, dot product, Euclidean, and Manhattan distances to find the best cutoff for calling a pair positive or negative, but the final score you see is always expressed in cosine-similarity terms. This matters because the threshold that gets selected may have been optimal under a different metric, so the reported average precision blends a multi-metric search with a single standardized reporting format.