imagining syntax

← schedopt_v4

fixed_3.0

done · peak unseen_mismatch 73.5

α schedule

α 3.0 ×400

direction
constant
α range
3.0–3.0
stages
1

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 3.0

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
100 54.8 ± 3.851.9 ± 2.651.3 ± 3.350.3 ± 1.9
110 57.7 ± 6.153.2 ± 4.353.1 ± 4.851.1 ± 2.7
120 60.9 ± 10.356.4 ± 7.156.2 ± 8.852.3 ± 5.1
130 64.3 ± 11.959.7 ± 9.758.8 ± 11.554.7 ± 8.1
140 70.6 ± 11.865.4 ± 9.963.7 ± 12.258.2 ± 7.9
150 76.8 ± 12.171.7 ± 10.768.9 ± 13.362.5 ± 8.8
160 80.8 ± 10.175.2 ± 9.072.1 ± 12.364.9 ± 9.5
170 85.2 ± 8.580.4 ± 7.475.0 ± 12.868.8 ± 9.9
180 86.7 ± 8.283.0 ± 7.076.1 ± 12.970.9 ± 10.1
190 88.4 ± 7.285.3 ± 6.377.3 ± 12.173.4 ± 10.1
200 87.9 ± 5.685.2 ± 5.775.0 ± 9.871.0 ± 8.0
210 88.2 ± 4.785.9 ± 4.774.7 ± 9.071.6 ± 7.5
220 88.8 ± 4.086.8 ± 3.975.2 ± 7.772.3 ± 6.9
230 89.1 ± 3.487.0 ± 3.274.9 ± 6.771.6 ± 5.9
240 89.6 ± 3.187.4 ± 3.174.7 ± 6.072.0 ± 5.6
250 89.7 ± 3.188.2 ± 2.875.8 ± 6.472.6 ± 6.4
260 90.3 ± 3.088.3 ± 2.575.6 ± 5.972.7 ± 4.8
270 90.1 ± 2.788.7 ± 2.275.5 ± 5.172.9 ± 4.5
280 90.8 ± 2.489.0 ± 1.975.7 ± 4.772.4 ± 3.9
290 91.0 ± 2.289.1 ± 1.975.5 ± 4.372.3 ± 4.1
300 91.3 ± 2.089.3 ± 1.976.7 ± 3.873.2 ± 2.9
310 91.2 ± 2.089.5 ± 1.776.6 ± 4.073.4 ± 4.1
320 91.3 ± 2.189.3 ± 1.976.5 ± 4.372.8 ± 3.4
330 91.4 ± 1.989.7 ± 2.076.4 ± 4.072.9 ± 3.4
340 92.0 ± 1.989.9 ± 2.176.0 ± 4.472.9 ± 3.9
350 92.2 ± 1.189.7 ± 2.076.0 ± 3.672.3 ± 4.6
360 92.2 ± 1.890.0 ± 1.677.5 ± 4.073.5 ± 4.1
370 92.2 ± 1.390.3 ± 2.077.1 ± 4.273.3 ± 4.3
380 91.9 ± 1.089.8 ± 1.575.7 ± 4.171.8 ± 4.1
390 92.3 ± 1.190.2 ± 1.876.2 ± 4.572.8 ± 4.4
400 92.3 ± 0.990.4 ± 1.376.3 ± 4.772.5 ± 4.4

Downloads