Table 35
1 Question2 Answer and tokens3 Hidden states4 Tag each token5 Check
OSafeNot part of a hallucination
B-HALLHallucination beginsThe first wrong token, the onset
I-HALLHallucination continuesThe later wrong tokens
The answer, token by token The tag under each token is the tagger's output.
token
its 20,480 numbers
BiLSTMreads the context
CRFchecks the tag order
Otag
Onset score of each token How strongly the tagger believes a hallucination begins at this token, from 0 to 1.
What the label says is wrong Marked before our experiment, never shown to the tagger.
Recorded demo: all questions and outputs come from an earlier saved run. No model runs live.