use case · multi-model
Fill-mask ranks candidates by how likely they are in context — not by what they mean. So compose two models: BERT generates the candidate words for the blank, then all-MiniLM-L6 embeds each completed sentence and re-ranks them by cosine similarity to a target sentence you choose. The result shows how a grammatically-likely fill can be semantically miles from what you meant — and both models run locally.
BERT fill-mask (~110 MB) proposes; MiniLM embeddings (~25 MB) ranks. Each auto-initialises from cache when present.
Each row shows BERT's fill probability (how likely the word is in the blank) and MiniLM's cosine similarity (how close the finished sentence is to your target). Sorting by cosine surfaces the fills that actually match your intent — a tiny pipeline you could drop into a controllable text tool.