imagining syntax

← schedopt_v5_confirm

free-f457fa6e

done · peak unseen_mismatch 97.2

α schedule

α 4.6 ×10 α 4.7 ×10 α 4.8 ×10 α 5.0 ×10 α 4.3 ×10 α 4.7 ×10 α 4.6 ×10 α 5.0 ×20 α 4.6 ×10 α 3.7 ×10 α 4.3 ×10 α 4.2 ×20 α 4.7 ×10 α 5.0 ×10 α 4.7 ×10 α 1.6 ×10 α 0.0 ×30 α 0.1 ×10 α 2.7 ×10 α 3.6 ×10 α 3.9 ×10 α 5.0 ×10 α 2.7 ×10 α 1.0 ×10 α 0.7 ×10 α 0.0 ×110

direction
mixed
α range
0.0–5.0
stages
26

Figures

schedule.png
schedule.png

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 0.0

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
100 60.8 ± 7.857.3 ± 6.154.1 ± 5.951.6 ± 3.6
110 65.6 ± 9.161.8 ± 7.857.0 ± 7.653.8 ± 5.3
120 70.6 ± 9.566.8 ± 8.559.6 ± 8.756.2 ± 6.5
130 74.9 ± 8.671.4 ± 7.962.1 ± 8.958.8 ± 6.9
140 77.4 ± 7.574.5 ± 7.063.6 ± 8.460.4 ± 6.5
150 79.1 ± 6.076.5 ± 5.364.1 ± 7.661.4 ± 5.9
160 79.9 ± 4.977.4 ± 4.564.5 ± 7.261.7 ± 5.6
170 80.3 ± 4.378.1 ± 3.563.9 ± 6.661.5 ± 5.2
180 83.8 ± 4.681.6 ± 4.268.6 ± 7.565.7 ± 6.4
190 89.3 ± 6.386.9 ± 6.479.5 ± 10.376.0 ± 9.9
200 94.9 ± 6.592.6 ± 7.491.4 ± 8.888.1 ± 9.6
210 97.6 ± 5.393.1 ± 10.796.5 ± 6.790.4 ± 12.2
220 98.5 ± 4.773.6 ± 25.197.9 ± 6.471.0 ± 25.6
230 99.2 ± 2.683.2 ± 19.798.5 ± 3.880.8 ± 19.2
240 99.2 ± 1.593.7 ± 10.998.5 ± 2.891.6 ± 11.3
250 96.5 ± 3.489.7 ± 0.992.9 ± 7.185.7 ± 1.3
260 94.7 ± 5.091.9 ± 3.990.0 ± 9.888.1 ± 7.5
270 91.1 ± 0.089.4 ± 0.079.3 ± 0.078.4 ± 0.0
280 95.5 ± 0.094.0 ± 0.086.1 ± 0.086.0 ± 0.0
290 99.3 ± 0.099.2 ± 0.097.5 ± 0.097.2 ± 0.0

Downloads