imagining syntax

← schedopt_v5_confirm

phi3star

done · peak unseen_mismatch 93.1

Referenced by

α schedule

α 3.1993 ×25 α 3.155 ×25 α 3.0753 ×25 α 2.963 ×25 α 2.8198 ×25 α 2.6469 ×25 α 2.4453 ×25 α 2.2158 ×25 α 1.9588 ×25 α 1.675 ×25 α 1.3649 ×25 α 1.0288 ×25 α 0.6671 ×25 α 0.2802 ×25 α 0.2802 ×50

direction
descending
α range
0.2802–3.1993
stages
15

Figures

schedule.png
schedule.png

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 0.2802

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
100 56.4 ± 6.652.8 ± 4.152.8 ± 5.850.5 ± 3.1
110 59.8 ± 9.155.5 ± 6.555.1 ± 8.351.9 ± 4.8
120 64.3 ± 11.259.7 ± 8.858.2 ± 10.554.3 ± 6.8
130 70.0 ± 12.464.8 ± 10.762.6 ± 12.457.4 ± 8.8
140 75.0 ± 12.869.8 ± 11.566.5 ± 13.761.1 ± 10.8
150 79.6 ± 11.974.8 ± 10.770.4 ± 13.765.0 ± 11.1
160 83.5 ± 11.179.3 ± 10.573.9 ± 13.469.0 ± 11.7
170 86.6 ± 9.382.7 ± 9.276.7 ± 12.671.9 ± 11.3
180 88.9 ± 7.585.5 ± 7.779.0 ± 11.274.3 ± 10.0
190 90.9 ± 6.387.9 ± 6.481.2 ± 10.076.6 ± 9.1
200 92.2 ± 5.389.6 ± 5.382.5 ± 8.878.2 ± 8.4
210 93.4 ± 4.790.9 ± 4.683.7 ± 8.279.6 ± 7.8
220 94.3 ± 4.092.2 ± 4.284.7 ± 7.380.8 ± 7.1
230 95.2 ± 3.593.3 ± 3.686.4 ± 6.782.6 ± 6.7
240 96.4 ± 2.594.7 ± 2.688.7 ± 5.985.2 ± 5.8
250 96.8 ± 2.195.2 ± 2.289.1 ± 4.685.5 ± 4.6
260 97.8 ± 1.796.6 ± 1.991.8 ± 4.188.3 ± 4.6
270 98.4 ± 1.597.2 ± 1.793.7 ± 3.589.6 ± 4.4
280 98.8 ± 1.297.5 ± 2.194.6 ± 2.989.8 ± 4.7
290 99.5 ± 0.598.4 ± 1.097.0 ± 1.792.1 ± 3.0
300 99.9 ± 0.399.0 ± 0.498.1 ± 1.693.1 ± 2.8
310 100.0 ± 0.098.7 ± 1.099.0 ± 0.288.9 ± 7.1
320 100.0 ± 0.098.7 ± 0.399.7 ± 0.185.5 ± 3.8
330 100.0 ± 0.098.9 ± 0.399.9 ± 0.189.2 ± 0.1
340 100.0 ± 0.199.0 ± 0.3100.0 ± 0.191.9 ± 0.7

Downloads