imagining syntax

← schedopt_v4

fixed_2.1

done · peak unseen_mismatch 87.3

α schedule

α 2.1 ×400

direction
constant
α range
2.1–2.1
stages
1

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 2.1

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
100 53.2 ± 6.750.1 ± 1.251.6 ± 6.549.7 ± 2.0
110 53.9 ± 7.650.2 ± 2.052.4 ± 7.650.2 ± 2.2
120 54.3 ± 10.452.3 ± 7.753.8 ± 10.451.5 ± 7.0
130 55.1 ± 10.552.4 ± 6.654.1 ± 10.651.4 ± 6.4
140 57.2 ± 11.253.7 ± 9.956.0 ± 10.552.3 ± 8.7
150 59.8 ± 12.555.4 ± 11.158.4 ± 11.553.2 ± 10.0
160 62.6 ± 12.555.6 ± 8.660.1 ± 11.253.3 ± 7.3
170 68.9 ± 15.560.5 ± 12.466.0 ± 16.057.8 ± 11.5
180 71.7 ± 16.462.8 ± 12.368.4 ± 17.059.5 ± 11.0
190 79.4 ± 16.369.4 ± 12.776.2 ± 17.065.7 ± 12.0
200 85.3 ± 14.874.1 ± 12.782.6 ± 15.770.2 ± 11.3
210 90.9 ± 12.679.2 ± 13.187.8 ± 14.475.8 ± 13.0
220 93.1 ± 10.282.3 ± 13.490.3 ± 13.077.9 ± 14.7
230 93.5 ± 10.281.9 ± 14.091.2 ± 12.178.5 ± 13.9
240 94.6 ± 6.881.7 ± 10.591.1 ± 10.677.4 ± 11.0
250 96.3 ± 5.183.8 ± 13.093.7 ± 7.879.1 ± 11.8
260 96.0 ± 5.088.5 ± 7.593.0 ± 7.583.9 ± 7.3
270 96.6 ± 3.091.5 ± 4.393.1 ± 5.686.3 ± 5.6
280 97.4 ± 2.092.1 ± 6.493.9 ± 4.387.3 ± 6.2
290 97.4 ± 1.592.6 ± 4.894.8 ± 2.587.3 ± 5.2
300 97.8 ± 1.191.8 ± 6.394.6 ± 1.584.6 ± 8.1
310 98.2 ± 0.991.2 ± 6.194.1 ± 0.683.0 ± 8.4
320 97.4 ± 1.591.9 ± 4.392.7 ± 3.583.0 ± 1.6
330 97.7 ± 1.695.4 ± 1.793.1 ± 3.786.7 ± 3.9
340 97.6 ± 1.795.5 ± 2.991.9 ± 4.887.0 ± 6.5
350 95.0 ± 0.193.0 ± 0.187.3 ± 1.982.2 ± 2.5
360 96.3 ± 0.194.8 ± 1.288.8 ± 2.583.8 ± 3.8
370 96.5 ± 1.195.2 ± 1.289.3 ± 0.883.5 ± 0.8
380 96.6 ± 1.495.6 ± 1.689.4 ± 2.683.3 ± 3.1
390 96.5 ± 0.394.8 ± 1.288.4 ± 1.583.0 ± 0.2
400 96.8 ± 0.795.0 ± 0.988.1 ± 1.383.2 ± 0.6

Downloads