vec2vec

vec2vec is the first method to successfully translate embeddings from the space of one model to another without any paired data. vec2vec translations achieve cosine similarity as high as 0.92 to the ground-truth vectors in their target embedding spaces and perfect matching on over 8000 shuffled embeddings (without access to the set of possible matches in advance).¹

Footnotes

1. Harnessing the Universal Geometry of Embeddings ↩