Sent to the LLM (0.2 ≤ P < 0.8)
IN THE BAND
Jev
Jev: P = 0.03, annotators split
Jev: P = 0.03, annotators split
Jev: P = 0.04, annotators split
Jev: P = 0.05, annotators split
Jev: P = 0.05, both annotators disagree
Jev: P = 0.06, annotators split
Jev: P = 0.08, annotators split
Jev: P = 0.09, both annotators disagree
Jev: P = 0.12, annotators split
Jev: P = 0.12, annotators split
Jev: P = 0.13, both annotators disagree
Jev: P = 0.16, annotators split
Jev: P = 0.22, both annotators disagree
Jev: P = 0.31, both annotators disagree
Jev: P = 0.33, both annotators disagree
Jev: P = 0.39, annotators split
Jev: P = 0.4, annotators split
Jev: P = 0.5, annotators split
Jev: P = 0.52, both annotators disagree
Jev: P = 0.56, both annotators disagree
Jev: P = 0.8, annotators split
5 of 8
clear mistakes caught
7.6% of queries sent
0
0.2
0.5
0.8
1
djev
djev: P = 1e-06, annotators split
djev: P = 6e-06, annotators split
djev: P = 1.1e-05, annotators split
djev: P = 1.6e-05, annotators split
djev: P = 0.000289, annotators split
djev: P = 0.000308, both annotators disagree
djev: P = 0.000625, annotators split
djev: P = 0.000642, annotators split
djev: P = 0.000806, annotators split
djev: P = 0.000923, annotators split
djev: P = 0.00429, both annotators disagree
djev: P = 0.00591, annotators split
djev: P = 0.00656, both annotators disagree
djev: P = 0.0179, annotators split
djev: P = 0.0194, both annotators disagree
djev: P = 0.022, annotators split
djev: P = 0.039, both annotators disagree
djev: P = 0.0442, both annotators disagree
djev: P = 0.0974, both annotators disagree
djev: P = 0.136, both annotators disagree
djev: P = 0.219, both annotators disagree
djev: P = 0.232, both annotators disagree
djev: P = 0.294, both annotators disagree
djev: P = 0.422, both annotators disagree
djev: P = 0.447, both annotators disagree
djev: P = 0.477, both annotators disagree
djev: P = 0.486, both annotators disagree
djev: P = 0.488, both annotators disagree
djev: P = 0.524, annotators split
8 of 16
clear mistakes caught
6.7% of queries sent
0
0.2
0.5
0.8
1
Luna
Luna: P = 0, annotators split
Luna: P = 0, annotators split
Luna: P = 0, annotators split
Luna: P = 0, annotators split
Luna: P = 0, annotators split
Luna: P = 0, both annotators disagree
Luna: P = 0, annotators split
Luna: P = 0, annotators split
Luna: P = 0, annotators split
Luna: P = 0, annotators split
Luna: P = 0.01, both annotators disagree
Luna: P = 0.01, annotators split
Luna: P = 0.02, both annotators disagree
Luna: P = 0.02, both annotators disagree
Luna: P = 0.02, both annotators disagree
Luna: P = 0.03, both annotators disagree
Luna: P = 0.04, annotators split
Luna: P = 0.06, both annotators disagree
Luna: P = 0.25, both annotators disagree
Luna: P = 0.28, annotators split
Luna: P = 0.47, both annotators disagree
Luna: P = 0.6, both annotators disagree
Luna: P = 0.94, annotators split
3 of 10
clear mistakes caught
4.2% of queries sent
0
0.2
0.5
0.8
1
12 mistakes at P ≤ 0.01
Each dot is a query the model got wrong at threshold 0.5, placed at the probability it gave.
Both annotators disagree with the model
Annotators split (half a mistake)