imagining syntax

← refine_v1

smooth_K12

done · peak unseen_mismatch 95.0

α schedule

α 2.9286 ×75 α 2.7857 ×75 α 2.6429 ×75 α 2.5 ×75 α 2.3571 ×75 α 2.2143 ×75 α 2.0714 ×75 α 1.9286 ×75 α 1.7857 ×75 α 1.6428 ×75 α 1.5 ×75 α 1.3571 ×75

direction
descending
α range
1.3571–2.9286
stages
12

Figures

schedule.png
schedule.png
trajectory.png
trajectory.png

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 1.3571

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
25 50.7 ± 1.649.8 ± 1.350.1 ± 1.550.5 ± 1.7
50 51.4 ± 1.549.5 ± 1.750.2 ± 1.750.2 ± 1.7
75 52.3 ± 2.149.8 ± 1.450.9 ± 1.849.4 ± 1.7
100 53.6 ± 3.449.8 ± 1.952.6 ± 3.349.9 ± 1.9
125 62.8 ± 12.751.6 ± 8.761.0 ± 12.950.4 ± 7.9
150 80.7 ± 16.864.8 ± 13.578.3 ± 17.962.7 ± 13.4
175 91.4 ± 13.079.5 ± 15.090.2 ± 13.777.6 ± 15.1
200 97.5 ± 6.586.0 ± 14.696.4 ± 7.983.6 ± 15.2
225 98.6 ± 3.192.5 ± 8.598.1 ± 3.790.6 ± 9.0
250 99.3 ± 1.197.2 ± 2.998.6 ± 1.995.0 ± 4.2
275 99.2 ± 1.197.7 ± 2.197.9 ± 2.694.7 ± 5.1
300 98.7 ± 1.496.9 ± 2.696.4 ± 3.892.3 ± 6.5
325 98.4 ± 1.696.2 ± 3.095.4 ± 4.390.5 ± 7.8
350 97.8 ± 1.795.9 ± 3.293.4 ± 5.587.9 ± 9.0
375 97.4 ± 1.895.3 ± 3.191.8 ± 6.386.6 ± 9.7
400 97.3 ± 1.695.2 ± 2.991.1 ± 6.084.7 ± 9.6
450 97.1 ± 1.494.4 ± 2.490.0 ± 5.384.4 ± 8.1
500 97.3 ± 1.195.1 ± 1.990.7 ± 4.484.4 ± 7.3
550 97.8 ± 0.795.2 ± 1.291.3 ± 3.583.1 ± 7.2
600 97.8 ± 1.195.2 ± 1.891.3 ± 3.983.0 ± 6.4
650 98.4 ± 0.895.9 ± 1.992.2 ± 3.383.4 ± 6.1
700 99.0 ± 0.796.4 ± 1.994.7 ± 2.985.4 ± 7.0
750 99.0 ± 0.796.8 ± 2.194.4 ± 2.885.1 ± 8.6
800 99.6 ± 0.397.1 ± 1.995.7 ± 2.085.2 ± 8.9
850 99.7 ± 0.397.3 ± 3.297.6 ± 1.487.7 ± 8.6
900 99.7 ± 0.497.5 ± 2.297.0 ± 1.486.7 ± 8.8

Downloads