imagining syntax

← refine_v1

smooth_K24

done · peak unseen_mismatch 90.1

α schedule

α 2.9643 ×38 α 2.8929 ×38 α 2.8214 ×38 α 2.75 ×38 α 2.6786 ×38 α 2.6071 ×38 α 2.5357 ×38 α 2.4643 ×38 α 2.3929 ×38 α 2.3214 ×38 α 2.25 ×38 α 2.1786 ×38 α 2.1071 ×37 α 2.0357 ×37 α 1.9643 ×37 α 1.8928 ×37 α 1.8214 ×37 α 1.75 ×37 α 1.6786 ×37 α 1.6071 ×37 α 1.5357 ×37 α 1.4643 ×37 α 1.3928 ×37 α 1.3214 ×37

direction
descending
α range
1.3214–2.9643
stages
24

Figures

schedule.png
schedule.png
trajectory.png
trajectory.png

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 1.3214

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
25 50.5 ± 2.049.5 ± 1.650.0 ± 1.749.8 ± 1.9
50 51.4 ± 1.850.0 ± 1.049.7 ± 2.249.8 ± 1.7
75 52.4 ± 2.249.3 ± 1.651.0 ± 1.949.0 ± 1.7
100 55.7 ± 5.849.6 ± 2.653.9 ± 5.448.7 ± 2.2
125 68.0 ± 16.655.9 ± 7.264.8 ± 17.752.5 ± 5.9
150 84.2 ± 16.370.6 ± 13.681.0 ± 18.266.6 ± 13.8
175 92.8 ± 10.184.0 ± 15.189.8 ± 14.180.3 ± 16.4
200 97.4 ± 4.591.2 ± 12.595.3 ± 7.487.7 ± 13.4
225 97.8 ± 3.294.3 ± 8.195.2 ± 6.490.1 ± 9.1
250 97.9 ± 2.393.7 ± 9.994.5 ± 5.688.5 ± 10.9
275 97.4 ± 2.695.5 ± 3.393.0 ± 6.789.2 ± 8.6
300 97.1 ± 2.594.8 ± 3.691.9 ± 5.587.3 ± 8.1
325 96.9 ± 2.094.8 ± 3.290.1 ± 6.685.2 ± 9.2
350 96.5 ± 2.394.1 ± 2.789.7 ± 7.084.6 ± 8.5
375 96.4 ± 1.994.0 ± 2.688.5 ± 6.082.9 ± 8.0
400 96.6 ± 1.894.6 ± 2.688.5 ± 6.682.8 ± 8.5
450 96.4 ± 1.493.7 ± 2.286.8 ± 5.279.9 ± 6.4
500 96.8 ± 1.094.8 ± 1.787.8 ± 4.480.6 ± 6.4
550 97.5 ± 0.895.1 ± 1.489.1 ± 4.181.3 ± 5.6
600 97.8 ± 0.695.8 ± 1.590.2 ± 2.782.1 ± 6.8
650 98.5 ± 0.596.3 ± 1.591.3 ± 2.682.4 ± 7.5
700 98.9 ± 0.496.8 ± 2.293.7 ± 2.683.5 ± 9.1
750 99.2 ± 0.396.9 ± 2.294.6 ± 2.684.0 ± 9.6
800 99.6 ± 0.297.2 ± 2.495.8 ± 2.184.6 ± 10.8
850 99.8 ± 0.297.3 ± 3.097.3 ± 1.583.5 ± 13.4
900 99.7 ± 0.396.8 ± 3.997.4 ± 1.982.3 ± 16.1

Downloads