imagining syntax

← faithful_v1

dedup_zipfian

done · peak unseen_mismatch 91.7

Configuration

n_layer
2
n_head
4
n_embd
256
base_seed
42

Figures

accuracy_vs_alpha.png
accuracy_vs_alpha.png

Results

seen_* generated at each grid point's own training α · unseen_* = held-out pairings, uniform (α-independent) · 1000 pairs per condition

αseen_matchseen_mismatchunseen_matchunseen_mismatch
0.0 100.0 ± 0.022.8 ± 21.2100.0 ± 0.022.1 ± 19.6
0.1 100.0 ± 0.021.0 ± 24.5100.0 ± 0.020.9 ± 24.7
0.2 100.0 ± 0.04.6 ± 12.8100.0 ± 0.04.6 ± 12.5
0.3 100.0 ± 0.010.8 ± 18.6100.0 ± 0.010.6 ± 18.0
0.4 100.0 ± 0.09.0 ± 14.8100.0 ± 0.09.3 ± 15.0
0.5 100.0 ± 0.043.2 ± 37.1100.0 ± 0.043.8 ± 37.4
0.6 100.0 ± 0.013.7 ± 19.7100.0 ± 0.013.9 ± 19.9
0.7 100.0 ± 0.09.6 ± 16.0100.0 ± 0.09.4 ± 16.0
0.8 100.0 ± 0.028.1 ± 26.8100.0 ± 0.027.4 ± 26.8
0.9 100.0 ± 0.033.7 ± 30.5100.0 ± 0.033.4 ± 30.6
1.0 100.0 ± 0.044.7 ± 37.6100.0 ± 0.044.0 ± 37.1
1.1 100.0 ± 0.059.2 ± 31.5100.0 ± 0.058.8 ± 31.8
1.2 100.0 ± 0.065.8 ± 28.9100.0 ± 0.064.0 ± 29.4
1.3 99.9 ± 0.191.4 ± 8.699.8 ± 0.487.6 ± 10.2
1.4 99.8 ± 0.491.4 ± 12.399.1 ± 1.486.1 ± 17.3
1.5 99.8 ± 0.296.8 ± 2.596.6 ± 3.290.1 ± 6.0
1.6 99.7 ± 0.298.4 ± 1.197.7 ± 2.191.7 ± 5.2
1.7 99.4 ± 0.498.1 ± 1.395.8 ± 2.189.2 ± 5.0
1.8 99.6 ± 0.598.9 ± 0.995.0 ± 2.288.2 ± 5.2
1.9 99.6 ± 0.397.5 ± 3.291.1 ± 2.680.8 ± 7.3
2.0 99.6 ± 0.398.1 ± 1.491.7 ± 3.080.9 ± 6.0
2.1 99.6 ± 0.199.0 ± 0.589.2 ± 2.182.4 ± 5.9
2.2 99.5 ± 0.499.2 ± 0.689.4 ± 2.681.7 ± 9.5
2.3 99.6 ± 0.399.0 ± 0.486.8 ± 3.477.7 ± 7.8
2.4 99.7 ± 0.299.1 ± 0.586.6 ± 2.676.2 ± 6.4
2.5 99.8 ± 0.199.5 ± 0.285.8 ± 3.678.1 ± 6.2
2.6 99.6 ± 0.299.4 ± 0.284.6 ± 3.374.8 ± 5.2
2.7 99.8 ± 0.199.5 ± 0.381.4 ± 3.673.0 ± 5.6
2.8 99.8 ± 0.199.6 ± 0.283.2 ± 3.475.8 ± 5.0
2.9 99.8 ± 0.299.7 ± 0.381.1 ± 3.672.1 ± 5.9
3.0 99.8 ± 0.299.8 ± 0.181.6 ± 2.771.9 ± 6.2

Downloads