imagining syntax

← refine_v1

smooth_K48

done · peak unseen_mismatch 91.2

α schedule

α 2.9821 ×19 α 2.9464 ×19 α 2.9107 ×19 α 2.875 ×19 α 2.8393 ×19 α 2.8036 ×19 α 2.7679 ×19 α 2.7321 ×19 α 2.6964 ×19 α 2.6607 ×19 α 2.625 ×19 α 2.5893 ×19 α 2.5536 ×19 α 2.5179 ×19 α 2.4821 ×19 α 2.4464 ×19 α 2.4107 ×19 α 2.375 ×19 α 2.3393 ×19 α 2.3036 ×19 α 2.2679 ×19 α 2.2321 ×19 α 2.1964 ×19 α 2.1607 ×19 α 2.125 ×19 α 2.0893 ×19 α 2.0536 ×19 α 2.0178 ×19 α 1.9821 ×19 α 1.9464 ×19 α 1.9107 ×19 α 1.875 ×19 α 1.8393 ×19 α 1.8036 ×19 α 1.7678 ×19 α 1.7321 ×19 α 1.6964 ×18 α 1.6607 ×18 α 1.625 ×18 α 1.5893 ×18 α 1.5536 ×18 α 1.5178 ×18 α 1.4821 ×18 α 1.4464 ×18 α 1.4107 ×18 α 1.375 ×18 α 1.3393 ×18 α 1.3036 ×18

direction
descending
α range
1.3036–2.9821
stages
48

Figures

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

Results

accuracy by training iteration, mean ± sd over seeds · schedule ends at α = 1.3036

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

iterationseen_matchseen_mismatchunseen_matchunseen_mismatch
25 50.9 ± 1.749.9 ± 1.550.3 ± 1.550.4 ± 2.4
50 50.9 ± 1.050.4 ± 1.449.5 ± 1.349.7 ± 1.6
75 50.8 ± 1.949.6 ± 1.250.5 ± 1.949.3 ± 1.6
100 54.9 ± 4.049.5 ± 1.851.6 ± 2.850.1 ± 2.0
125 62.0 ± 9.553.8 ± 4.758.6 ± 10.251.2 ± 4.0
150 83.7 ± 13.567.2 ± 10.880.2 ± 15.463.4 ± 10.8
175 95.7 ± 4.684.1 ± 11.794.0 ± 6.380.6 ± 12.5
200 99.0 ± 1.089.5 ± 12.598.0 ± 2.087.7 ± 11.4
225 98.6 ± 1.694.2 ± 9.697.4 ± 3.191.2 ± 9.9
250 98.5 ± 1.294.9 ± 6.996.4 ± 3.291.0 ± 7.7
275 98.0 ± 1.696.1 ± 2.894.8 ± 3.890.9 ± 5.6
300 97.3 ± 1.895.1 ± 2.793.0 ± 4.988.0 ± 6.0
325 96.7 ± 2.094.4 ± 2.891.2 ± 5.185.3 ± 6.9
350 96.6 ± 1.594.3 ± 2.690.0 ± 5.183.9 ± 6.8
375 96.0 ± 1.793.5 ± 2.488.0 ± 5.681.7 ± 6.9
400 96.4 ± 1.394.1 ± 2.388.6 ± 4.582.0 ± 6.4
450 96.5 ± 1.494.2 ± 2.088.6 ± 3.680.9 ± 5.7
500 97.1 ± 1.094.7 ± 1.788.8 ± 3.180.6 ± 5.6
550 97.1 ± 1.094.7 ± 1.489.4 ± 2.580.6 ± 3.9
600 97.9 ± 0.995.7 ± 2.090.4 ± 3.281.3 ± 6.2
650 98.5 ± 0.896.0 ± 1.992.4 ± 2.282.2 ± 6.7
700 98.8 ± 0.396.2 ± 1.693.3 ± 2.381.3 ± 6.6
750 99.3 ± 0.496.6 ± 1.794.3 ± 2.081.3 ± 7.6
800 99.6 ± 0.297.2 ± 1.595.1 ± 3.082.4 ± 7.8
850 99.7 ± 0.297.3 ± 2.096.7 ± 1.783.0 ± 10.4
900 99.8 ± 0.297.5 ± 3.497.2 ± 2.083.2 ± 11.8

Downloads