imagining syntax

← schedopt_v5_confirm

phi4star

done · peak unseen_mismatch 94.5

Referenced by

α schedule

α 5.0 ×25 α 5.0 ×25 α 5.0 ×25 α 1.0634 ×25 α 1.0536 ×25 α 5.0 ×25 α 5.0 ×25 α 5.0 ×25 α 0.9043 ×25 α 0.8215 ×25 α 0.7122 ×25 α 3.084 ×25 α 2.1282 ×25 α 0.1755 ×25 α 0.1755 ×50

direction
mixed
α range
0.1755–5.0
stages
15

Figures

schedule.png
schedule.png

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 0.1755

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 ± 8.252.4 ± 4.453.8 ± 7.650.4 ± 3.4
110 59.7 ± 11.653.0 ± 5.856.8 ± 11.251.2 ± 4.4
120 64.0 ± 15.254.4 ± 7.161.3 ± 15.352.6 ± 6.1
130 70.6 ± 17.557.1 ± 10.467.8 ± 18.154.6 ± 9.3
140 77.1 ± 17.564.3 ± 13.673.7 ± 18.460.8 ± 13.0
150 82.2 ± 16.272.1 ± 14.478.3 ± 17.768.0 ± 14.3
160 85.5 ± 13.778.2 ± 12.980.8 ± 15.772.7 ± 13.6
170 87.6 ± 11.482.1 ± 11.181.9 ± 13.975.5 ± 12.4
180 88.5 ± 9.084.1 ± 8.981.8 ± 11.676.5 ± 10.8
190 88.8 ± 7.885.6 ± 7.581.7 ± 10.576.9 ± 9.8
200 88.4 ± 6.685.8 ± 6.479.7 ± 9.875.7 ± 8.8
210 92.7 ± 7.090.1 ± 7.387.3 ± 10.283.4 ± 10.3
220 96.0 ± 7.293.5 ± 8.193.3 ± 9.789.5 ± 10.1
230 95.6 ± 9.491.8 ± 10.693.5 ± 11.088.3 ± 12.1
240 93.7 ± 11.885.2 ± 12.892.4 ± 13.181.8 ± 12.6
250 95.1 ± 11.080.8 ± 14.694.0 ± 12.778.0 ± 14.1
260 95.5 ± 11.179.9 ± 22.794.6 ± 13.876.7 ± 21.7
270 95.4 ± 9.966.8 ± 27.993.8 ± 13.666.6 ± 26.9
280 96.3 ± 7.371.8 ± 35.295.6 ± 8.870.2 ± 34.0
290 98.7 ± 1.872.6 ± 26.198.1 ± 2.770.2 ± 24.9
300 98.3 ± 1.787.4 ± 0.898.0 ± 1.987.7 ± 2.6
310 98.7 ± 1.393.7 ± 1.798.0 ± 2.092.9 ± 2.2
320 98.2 ± 0.095.9 ± 0.097.0 ± 0.094.5 ± 0.0

Downloads