imagining syntax

← schedopt_v4

free-3dd752ef

done · peak unseen_mismatch 87.5

α schedule

α 3.558 ×34 α 1.4 ×12 α 4.617 ×83 α 0.7674 ×14 α 2.912 ×11 α 3.64 ×11 α 3.826 ×18 α 0.3263 ×17 α 0.272 ×39 α 1.941 ×25 α 2.966 ×36 α 2.966 ×100

direction
mixed
α range
0.272–4.617
stages
12

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 2.966

seen_* probed at reference α = 1.4 · unseen_* = held-out pairings, uniform (α-independent) · 1000 pairs per condition

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
100 57.0 ± 8.752.7 ± 5.253.6 ± 7.750.6 ± 3.7
110 61.0 ± 11.056.0 ± 7.055.9 ± 10.252.1 ± 4.7
120 65.5 ± 12.960.3 ± 9.458.8 ± 12.054.6 ± 6.7
130 70.6 ± 12.865.7 ± 10.962.0 ± 12.557.7 ± 8.9
140 73.0 ± 14.667.4 ± 12.566.8 ± 14.660.7 ± 11.3
150 79.7 ± 16.273.6 ± 15.074.1 ± 17.567.1 ± 14.8
160 83.0 ± 14.677.5 ± 13.876.3 ± 16.170.2 ± 13.6
170 85.4 ± 11.781.1 ± 11.477.4 ± 13.872.0 ± 12.3
180 86.6 ± 9.582.9 ± 9.176.8 ± 12.572.3 ± 10.8
190 89.1 ± 10.085.3 ± 10.081.8 ± 12.776.3 ± 11.6
200 93.2 ± 11.388.2 ± 13.490.4 ± 13.184.0 ± 15.0
210 93.8 ± 12.282.8 ± 17.192.4 ± 13.379.7 ± 18.2
220 91.4 ± 15.860.3 ± 22.290.5 ± 15.958.3 ± 20.9
230 94.6 ± 13.263.6 ± 16.994.2 ± 13.461.8 ± 16.8
240 95.8 ± 13.060.1 ± 16.395.7 ± 13.259.3 ± 15.9
250 96.5 ± 11.761.3 ± 19.996.0 ± 13.260.4 ± 19.7
260 96.1 ± 12.267.3 ± 20.295.6 ± 14.065.5 ± 19.3
270 96.1 ± 11.777.2 ± 14.095.7 ± 13.175.0 ± 14.8
280 96.3 ± 9.883.3 ± 15.494.9 ± 13.180.6 ± 16.9
290 95.7 ± 8.382.3 ± 17.393.0 ± 12.978.3 ± 17.9
300 95.6 ± 7.281.9 ± 15.891.4 ± 12.776.7 ± 15.9
310 95.6 ± 5.777.5 ± 14.190.1 ± 11.272.3 ± 11.0
320 94.3 ± 5.480.2 ± 9.788.8 ± 9.973.4 ± 7.3
330 94.9 ± 5.685.5 ± 4.489.6 ± 10.978.5 ± 2.1
340 95.3 ± 4.192.5 ± 5.491.7 ± 6.787.5 ± 9.0
350 94.0 ± 3.190.8 ± 1.687.6 ± 4.783.9 ± 0.5
360 91.9 ± 4.287.3 ± 2.082.5 ± 5.278.6 ± 0.3
370 92.9 ± 2.987.7 ± 0.984.3 ± 3.878.2 ± 1.8
380 92.7 ± 1.987.7 ± 0.684.3 ± 2.878.4 ± 2.0
390 91.0 ± 1.188.0 ± 1.482.1 ± 0.977.2 ± 0.9
400 92.3 ± 1.189.1 ± 0.581.5 ± 0.378.0 ± 0.9

Downloads