mteb-glove-baseline-average-42

IN premise — summaries/2026/08/24/muennighoff-2022-mteb-s4-additional-modalities-text-embeddings-are-chunk-1.md

Created 2026-08-25T02:58:20+00:00

GloVe average-of-word-embeddings baseline scores approximately 42 average across MTEB tasks, establishing a floor for learned embeddings.

Summary

If you just average the pre-trained word vectors in a sentence, you already hit around 42 on the MTEB benchmark, so any serious embedding model needs to clear that bar to prove it is doing more than trivial word-averaging. This gives the system a concrete sanity floor: anything scoring below roughly 42 is underperforming compared to the simplest possible approach.