imagining syntax

← bimodality_grid_v2

alpha_grid

done · peak unseen_mismatch 85.0

Configuration

n_layer
2
n_head
4
n_embd
256

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
0.8 100.0 ± 0.035.4 ± 30.2100.0 ± 0.035.3 ± 29.8
1.0 100.0 ± 0.060.9 ± 32.4100.0 ± 0.060.4 ± 32.5
1.05 100.0 ± 0.063.9 ± 37.5100.0 ± 0.063.1 ± 37.1
1.1 100.0 ± 0.078.1 ± 26.1100.0 ± 0.177.1 ± 26.0
1.125 100.0 ± 0.180.9 ± 25.799.8 ± 0.579.4 ± 25.1
1.15 100.0 ± 0.185.8 ± 17.899.9 ± 0.384.0 ± 17.7
1.175 100.0 ± 0.187.7 ± 19.899.7 ± 0.885.0 ± 19.8
1.2 99.9 ± 0.283.3 ± 22.799.7 ± 0.780.9 ± 21.4

Downloads