use case · basics
The simplest thing masked language modelling does: hide one word and rank what belongs there.
Pick an example or write your own with a single [MASK], and watch the ranked list —
with the probability and the raw logit for each guess — update as you type.
Predictions
The bar is the softmax probability; the logit is the raw score the network produced
before normalising. On a strongly-constrained slot one guess towers over the rest; on an open slot
the probability spreads thin — a direct read on how sure the model is.
← Back to BERT fill-mask · Next: Practical →