vec2vec-data-efficiency-50k-near-1m

Summary

You can map one embedding space to another using just 50,000 training pairs and land almost as close to the target (cosine 0.74) as you would with a full million pairs (cosine 0.75), on a standard natural-questions evaluation. In practical terms, this means the mapping saturates fast: spending twenty times more data buys you only a negligible bump in quality, so the system can rely on small, cheap training sets for this kind of cross-space translation.