imagining syntax

← schedopt_v4

free-74bd9917

done · peak unseen_mismatch 82.3

α schedule

α 3.344 ×31 α 1.656 ×16 α 3.38 ×19 α 2.01 ×28 α 3.994 ×11 α 1.519 ×23 α 3.674 ×12 α 1.534 ×16 α 3.657 ×17 α 1.472 ×34 α 3.669 ×15 α 1.964 ×29 α 3.96 ×49 α 3.96 ×100

direction
mixed
α range
1.472–3.994
stages
14

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 3.96

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
100 54.1 ± 4.650.4 ± 2.252.2 ± 4.049.5 ± 1.8
110 56.7 ± 6.851.5 ± 3.153.9 ± 6.749.9 ± 2.5
120 59.5 ± 10.452.6 ± 5.156.1 ± 10.250.6 ± 4.1
130 64.3 ± 13.454.7 ± 6.760.5 ± 13.452.1 ± 5.6
140 70.9 ± 16.259.6 ± 9.866.5 ± 16.355.8 ± 8.7
150 75.8 ± 17.463.9 ± 11.272.2 ± 17.959.7 ± 9.9
160 81.9 ± 16.869.0 ± 13.678.6 ± 18.165.0 ± 13.1
170 85.9 ± 15.975.8 ± 14.882.2 ± 17.371.2 ± 14.5
180 88.2 ± 15.080.3 ± 14.984.5 ± 16.275.7 ± 14.7
190 90.7 ± 13.982.3 ± 14.787.7 ± 15.678.1 ± 14.3
200 91.2 ± 13.581.6 ± 14.188.8 ± 15.177.9 ± 13.7
210 92.8 ± 12.380.9 ± 15.390.3 ± 13.977.6 ± 14.7
220 93.3 ± 11.583.3 ± 13.489.3 ± 13.278.8 ± 13.1
230 93.3 ± 10.487.9 ± 11.189.2 ± 13.082.3 ± 12.7
240 93.0 ± 11.386.9 ± 13.489.3 ± 12.182.0 ± 12.3
250 93.7 ± 8.786.0 ± 12.088.9 ± 10.679.7 ± 10.1
260 94.0 ± 6.290.0 ± 7.188.0 ± 8.082.3 ± 7.5
270 92.5 ± 4.089.2 ± 4.484.4 ± 7.178.6 ± 6.9
280 92.4 ± 3.489.3 ± 3.984.3 ± 6.278.7 ± 7.7
290 92.1 ± 2.289.6 ± 2.383.2 ± 4.977.4 ± 6.7
300 91.7 ± 2.388.9 ± 1.582.0 ± 5.176.6 ± 6.8
310 91.5 ± 2.088.8 ± 2.481.5 ± 5.476.2 ± 7.3
320 91.4 ± 2.288.7 ± 2.581.4 ± 5.777.0 ± 6.9
330 91.5 ± 2.188.9 ± 2.780.6 ± 5.475.6 ± 7.5
340 90.3 ± 1.288.1 ± 2.678.6 ± 5.073.7 ± 7.5
350 90.7 ± 2.488.4 ± 2.679.7 ± 5.975.3 ± 7.1
360 91.0 ± 2.288.4 ± 2.978.5 ± 5.374.1 ± 7.4
370 90.5 ± 2.188.4 ± 3.078.7 ± 6.173.3 ± 7.5
380 90.1 ± 2.087.8 ± 3.078.0 ± 5.673.0 ± 7.1
390 90.5 ± 1.587.9 ± 2.778.5 ± 4.673.1 ± 7.1
400 90.6 ± 1.888.2 ± 3.178.3 ± 4.772.6 ± 7.8

Downloads