use case · practical
Two things a developer would actually ship. Autocomplete appends a mask to the end of your text and suggests the next word — accept one and keep going, entirely offline. Knowledge probing turns a fact into a fill-in-the-blank and reads out what the model "knows" — and, just as usefully, where it's unsure.
Type, then let BERT propose the next word. Click a suggestion to append it, and it'll suggest the word after that — a private, on-device compose assist.
Suggestions (click to append):
Each template hides the key fact behind [MASK]. A confident, correct
top guess means the fact survived pre-training; a spread-out list means it didn't — a quick,
honest way to audit what a small model actually encodes.
Probing is a real research technique (the "LAMA" probe) — but treat a small uncased BERT as a fuzzy memory, not a fact source: it happily fills capitals it saw often and fumbles rare ones. That gap is exactly what makes probing a useful audit.