use case · wild
Give the model a list and it finds the sentence that belongs the least — the one whose average similarity to everything else is lowest. No categories, no rules: just geometry. It also groups the rest into loose clusters and plots them all on a 2D map, so you can see the outlier float away.
Odd one out:
"Belonging" here is each sentence's mean cosine similarity to all the others. The lowest scorer is the outlier. Clustering is a tiny greedy pass over the same similarities — enough to colour the map, not a production k-means. It all runs on the 384-d vectors the model just produced.
← Back to MiniLM · Next: Multi-model →