Adversarial robustness for embedding & semantic-search models — paraphrase, hard-negatives, negation, out-of-domain calibration. Ranked by quality, or by efficiency (quality per MB) where tiny on-device models shine.
#
MODEL
COMPOSITE
ACC
PARA
HARD
NEG
CALIB
SIZE
SPEED
QUAL/MB
loading…
Run your own model: /embed-eval.js — BYOM, your model never leaves your machine.
a 270MB server-class embedder scores ~97 here. quality is one axis; quality-per-MB is the other.