imagining syntax

← schedopt_v4

fixed_2.9

done · peak unseen_mismatch 75.3

α schedule

α 2.9 ×400

direction
constant
α range
2.9–2.9
stages
1

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 2.9

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
100 53.7 ± 4.651.5 ± 2.051.4 ± 4.150.4 ± 2.0
110 56.6 ± 6.452.9 ± 3.853.0 ± 5.351.1 ± 2.7
120 59.8 ± 10.556.0 ± 7.355.2 ± 9.152.5 ± 5.5
130 62.5 ± 12.059.2 ± 9.658.5 ± 11.154.8 ± 7.8
140 67.4 ± 13.463.2 ± 10.561.7 ± 13.257.8 ± 8.4
150 72.2 ± 14.668.3 ± 13.665.7 ± 14.561.4 ± 11.2
160 77.9 ± 12.672.1 ± 13.869.8 ± 13.864.3 ± 11.8
170 81.5 ± 9.476.0 ± 9.672.0 ± 12.066.4 ± 9.3
180 83.9 ± 9.479.3 ± 8.573.2 ± 12.067.7 ± 8.6
190 86.8 ± 9.282.8 ± 8.074.8 ± 12.370.7 ± 9.8
200 87.8 ± 6.984.8 ± 6.775.9 ± 11.671.8 ± 9.5
210 89.0 ± 6.886.1 ± 6.677.1 ± 11.173.5 ± 9.8
220 89.3 ± 5.587.2 ± 5.276.6 ± 9.573.7 ± 8.7
230 89.7 ± 3.687.8 ± 3.876.1 ± 7.073.0 ± 6.3
240 89.8 ± 3.388.1 ± 3.375.4 ± 6.273.4 ± 6.1
250 89.7 ± 3.188.4 ± 3.075.6 ± 5.872.9 ± 5.5
260 90.5 ± 3.088.9 ± 2.976.7 ± 5.774.0 ± 5.4
270 90.6 ± 2.688.8 ± 2.576.1 ± 4.473.6 ± 4.0
280 91.3 ± 2.389.1 ± 3.076.6 ± 4.674.1 ± 4.0
290 91.5 ± 2.389.5 ± 2.276.1 ± 4.373.5 ± 3.8
300 91.6 ± 2.189.9 ± 2.577.7 ± 3.274.7 ± 3.0
310 91.8 ± 1.989.7 ± 2.476.9 ± 3.974.7 ± 3.5
320 91.5 ± 1.789.8 ± 2.076.5 ± 4.673.8 ± 4.0
330 92.2 ± 1.290.4 ± 1.376.8 ± 2.973.9 ± 2.1
340 92.6 ± 1.490.7 ± 1.576.9 ± 3.873.8 ± 2.9
350 92.4 ± 1.590.6 ± 1.677.4 ± 4.473.6 ± 4.5
360 92.8 ± 1.391.0 ± 1.479.0 ± 3.375.3 ± 3.6
370 92.9 ± 1.391.1 ± 1.378.3 ± 3.774.9 ± 4.1
380 92.2 ± 1.090.5 ± 1.476.6 ± 3.073.2 ± 3.6
390 92.5 ± 1.090.8 ± 1.377.2 ± 4.073.6 ± 3.7
400 92.6 ± 1.390.8 ± 1.377.2 ± 3.773.4 ± 3.1

Downloads