Create cv_sweep_mha_output.txt
Browse files- cv_sweep_mha_output.txt +1413 -0
cv_sweep_mha_output.txt
ADDED
|
@@ -0,0 +1,1413 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
MHA CV Relational Test — Prototype
|
| 2 |
+
Band: 0.13 < CV < 0.30
|
| 3 |
+
|
| 4 |
+
======================================================================
|
| 5 |
+
D=64 heads=1 head_dim=64
|
| 6 |
+
======================================================================
|
| 7 |
+
|
| 8 |
+
[pre-train]
|
| 9 |
+
emb : *0.1585*
|
| 10 |
+
Q_full : *0.1350*
|
| 11 |
+
K_full : *0.1313*
|
| 12 |
+
V_full : 0.1228
|
| 13 |
+
Q_h0 : *0.1313*
|
| 14 |
+
K_h0 : *0.1311*
|
| 15 |
+
V_h0 : *0.1399*
|
| 16 |
+
out_proj : 0.1123
|
| 17 |
+
cls_head : 0.0680
|
| 18 |
+
act_Q_h0 : 0.3437
|
| 19 |
+
act_K_h0 : 0.3364
|
| 20 |
+
act_V_h0 : 0.3227
|
| 21 |
+
act_post_h0 : *0.2838*
|
| 22 |
+
act_emb : *0.2821*
|
| 23 |
+
act_post_full : 0.3208
|
| 24 |
+
|
| 25 |
+
[epoch 25] loss=1.2333 acc=96.80%
|
| 26 |
+
emb : *0.1588*
|
| 27 |
+
Q_full : *0.1345*
|
| 28 |
+
K_full : *0.1327*
|
| 29 |
+
V_full : *0.1322*
|
| 30 |
+
Q_h0 : 0.1242
|
| 31 |
+
K_h0 : *0.1313*
|
| 32 |
+
V_h0 : 0.1238
|
| 33 |
+
out_proj : *0.1365*
|
| 34 |
+
cls_head : 0.1026
|
| 35 |
+
act_Q_h0 : 0.3933
|
| 36 |
+
act_K_h0 : 0.3881
|
| 37 |
+
act_V_h0 : 0.3043
|
| 38 |
+
act_post_h0 : *0.2560*
|
| 39 |
+
act_emb : *0.2868*
|
| 40 |
+
act_post_full : *0.2549*
|
| 41 |
+
|
| 42 |
+
[post-train] loss=0.3196 acc=98.80%
|
| 43 |
+
emb : *0.1549* Δ=-0.0035
|
| 44 |
+
Q_full : *0.1352* Δ=+0.0001
|
| 45 |
+
K_full : *0.1453* Δ=+0.0139
|
| 46 |
+
V_full : 0.1219 Δ=-0.0009
|
| 47 |
+
Q_h0 : 0.1268 Δ=-0.0045
|
| 48 |
+
K_h0 : *0.1374* Δ=+0.0063
|
| 49 |
+
V_h0 : 0.1263 Δ=-0.0136
|
| 50 |
+
out_proj : *0.1486* Δ=+0.0362
|
| 51 |
+
cls_head : 0.0975 Δ=+0.0295
|
| 52 |
+
act_Q_h0 : 0.3235 Δ=-0.0201
|
| 53 |
+
act_K_h0 : 0.4174 Δ=+0.0810
|
| 54 |
+
act_V_h0 : 0.4063 Δ=+0.0836
|
| 55 |
+
act_post_h0 : 0.3038 Δ=+0.0199
|
| 56 |
+
act_emb : *0.2970* Δ=+0.0149
|
| 57 |
+
act_post_full : 0.3234 Δ=+0.0026
|
| 58 |
+
|
| 59 |
+
======================================================================
|
| 60 |
+
D=64 heads=2 head_dim=32
|
| 61 |
+
======================================================================
|
| 62 |
+
|
| 63 |
+
[pre-train]
|
| 64 |
+
emb : *0.1585*
|
| 65 |
+
Q_full : *0.1350*
|
| 66 |
+
K_full : *0.1313*
|
| 67 |
+
V_full : 0.1228
|
| 68 |
+
Q_h0 : 0.1180
|
| 69 |
+
K_h0 : 0.1152
|
| 70 |
+
V_h0 : *0.1331*
|
| 71 |
+
Q_h1 : 0.1260
|
| 72 |
+
K_h1 : *0.1424*
|
| 73 |
+
V_h1 : 0.1086
|
| 74 |
+
out_proj : 0.1293
|
| 75 |
+
cls_head : 0.0703
|
| 76 |
+
act_Q_h0 : 0.3661
|
| 77 |
+
act_K_h0 : 0.3481
|
| 78 |
+
act_V_h0 : 0.3632
|
| 79 |
+
act_post_h0 : 0.3995
|
| 80 |
+
act_Q_h1 : 0.3431
|
| 81 |
+
act_K_h1 : 0.3838
|
| 82 |
+
act_V_h1 : 0.3669
|
| 83 |
+
act_post_h1 : 0.4001
|
| 84 |
+
act_emb : *0.2314*
|
| 85 |
+
act_post_full : *0.2782*
|
| 86 |
+
|
| 87 |
+
[epoch 25] loss=1.2451 acc=97.80%
|
| 88 |
+
emb : *0.1592*
|
| 89 |
+
Q_full : *0.1306*
|
| 90 |
+
K_full : *0.1376*
|
| 91 |
+
V_full : *0.1304*
|
| 92 |
+
Q_h0 : *0.1305*
|
| 93 |
+
K_h0 : 0.1205
|
| 94 |
+
V_h0 : 0.1248
|
| 95 |
+
Q_h1 : 0.1289
|
| 96 |
+
K_h1 : *0.1314*
|
| 97 |
+
V_h1 : 0.1205
|
| 98 |
+
out_proj : 0.1271
|
| 99 |
+
cls_head : 0.1026
|
| 100 |
+
act_Q_h0 : 0.4112
|
| 101 |
+
act_K_h0 : 0.3448
|
| 102 |
+
act_V_h0 : 0.3660
|
| 103 |
+
act_post_h0 : 0.3506
|
| 104 |
+
act_Q_h1 : 0.3511
|
| 105 |
+
act_K_h1 : 0.3286
|
| 106 |
+
act_V_h1 : 0.4482
|
| 107 |
+
act_post_h1 : 0.3415
|
| 108 |
+
act_emb : 0.3138
|
| 109 |
+
act_post_full : *0.2861*
|
| 110 |
+
|
| 111 |
+
[post-train] loss=0.3067 acc=99.20%
|
| 112 |
+
emb : *0.1606* Δ=+0.0021
|
| 113 |
+
Q_full : 0.1187 Δ=-0.0163
|
| 114 |
+
K_full : *0.1422* Δ=+0.0108
|
| 115 |
+
V_full : 0.1286 Δ=+0.0058
|
| 116 |
+
Q_h0 : 0.1244 Δ=+0.0065
|
| 117 |
+
K_h0 : *0.1363* Δ=+0.0211
|
| 118 |
+
V_h0 : *0.1412* Δ=+0.0081
|
| 119 |
+
Q_h1 : 0.1234 Δ=-0.0026
|
| 120 |
+
K_h1 : *0.1420* Δ=-0.0004
|
| 121 |
+
V_h1 : *0.1333* Δ=+0.0247
|
| 122 |
+
out_proj : *0.1468* Δ=+0.0175
|
| 123 |
+
cls_head : 0.0908 Δ=+0.0205
|
| 124 |
+
act_Q_h0 : 0.3814 Δ=+0.0153
|
| 125 |
+
act_K_h0 : 0.4524 Δ=+0.1042
|
| 126 |
+
act_V_h0 : 0.4259 Δ=+0.0627
|
| 127 |
+
act_post_h0 : 0.3327 Δ=-0.0669
|
| 128 |
+
act_Q_h1 : 0.4036 Δ=+0.0605
|
| 129 |
+
act_K_h1 : 0.4540 Δ=+0.0701
|
| 130 |
+
act_V_h1 : 0.4934 Δ=+0.1265
|
| 131 |
+
act_post_h1 : 0.3545 Δ=-0.0456
|
| 132 |
+
act_emb : *0.2472* Δ=+0.0158
|
| 133 |
+
act_post_full : *0.2776* Δ=-0.0006
|
| 134 |
+
|
| 135 |
+
======================================================================
|
| 136 |
+
D=64 heads=4 head_dim=16
|
| 137 |
+
======================================================================
|
| 138 |
+
|
| 139 |
+
[pre-train]
|
| 140 |
+
emb : *0.1585*
|
| 141 |
+
Q_full : *0.1350*
|
| 142 |
+
K_full : *0.1313*
|
| 143 |
+
V_full : 0.1228
|
| 144 |
+
Q_h0 : 0.0998
|
| 145 |
+
K_h0 : 0.1173
|
| 146 |
+
V_h0 : *0.1446*
|
| 147 |
+
Q_h1 : 0.1092
|
| 148 |
+
K_h1 : 0.1076
|
| 149 |
+
V_h1 : 0.1237
|
| 150 |
+
Q_h2 : 0.1096
|
| 151 |
+
K_h2 : 0.1164
|
| 152 |
+
V_h2 : 0.0922
|
| 153 |
+
Q_h3 : *0.1306*
|
| 154 |
+
K_h3 : 0.1286
|
| 155 |
+
V_h3 : 0.1043
|
| 156 |
+
out_proj : 0.1263
|
| 157 |
+
cls_head : 0.0687
|
| 158 |
+
act_Q_h0 : 0.4522
|
| 159 |
+
act_K_h0 : 0.4437
|
| 160 |
+
act_V_h0 : 0.5482
|
| 161 |
+
act_post_h0 : 0.4707
|
| 162 |
+
act_Q_h1 : 0.4953
|
| 163 |
+
act_K_h1 : 0.5016
|
| 164 |
+
act_V_h1 : 0.3824
|
| 165 |
+
act_post_h1 : 0.4986
|
| 166 |
+
act_Q_h2 : 0.4539
|
| 167 |
+
act_K_h2 : 0.4482
|
| 168 |
+
act_V_h2 : 0.4898
|
| 169 |
+
act_post_h2 : 0.4971
|
| 170 |
+
act_Q_h3 : 0.4894
|
| 171 |
+
act_K_h3 : 0.4595
|
| 172 |
+
act_V_h3 : 0.4272
|
| 173 |
+
act_post_h3 : 0.4833
|
| 174 |
+
act_emb : *0.2194*
|
| 175 |
+
act_post_full : 0.3495
|
| 176 |
+
|
| 177 |
+
[epoch 25] loss=1.2531 acc=97.60%
|
| 178 |
+
emb : *0.1679*
|
| 179 |
+
Q_full : 0.1181
|
| 180 |
+
K_full : 0.1263
|
| 181 |
+
V_full : *0.1367*
|
| 182 |
+
Q_h0 : 0.1098
|
| 183 |
+
K_h0 : 0.1103
|
| 184 |
+
V_h0 : *0.1683*
|
| 185 |
+
Q_h1 : 0.1075
|
| 186 |
+
K_h1 : 0.1100
|
| 187 |
+
V_h1 : 0.0915
|
| 188 |
+
Q_h2 : 0.0976
|
| 189 |
+
K_h2 : 0.1104
|
| 190 |
+
V_h2 : 0.0889
|
| 191 |
+
Q_h3 : 0.1273
|
| 192 |
+
K_h3 : *0.1371*
|
| 193 |
+
V_h3 : 0.1202
|
| 194 |
+
out_proj : *0.1366*
|
| 195 |
+
cls_head : 0.0992
|
| 196 |
+
act_Q_h0 : 0.4341
|
| 197 |
+
act_K_h0 : 0.5047
|
| 198 |
+
act_V_h0 : 0.4410
|
| 199 |
+
act_post_h0 : 0.4073
|
| 200 |
+
act_Q_h1 : 0.4890
|
| 201 |
+
act_K_h1 : 0.4802
|
| 202 |
+
act_V_h1 : 0.4857
|
| 203 |
+
act_post_h1 : 0.4622
|
| 204 |
+
act_Q_h2 : 0.5123
|
| 205 |
+
act_K_h2 : 0.4912
|
| 206 |
+
act_V_h2 : 0.4541
|
| 207 |
+
act_post_h2 : 0.4602
|
| 208 |
+
act_Q_h3 : 0.4813
|
| 209 |
+
act_K_h3 : 0.4718
|
| 210 |
+
act_V_h3 : 0.4958
|
| 211 |
+
act_post_h3 : 0.4196
|
| 212 |
+
act_emb : *0.2642*
|
| 213 |
+
act_post_full : *0.2649*
|
| 214 |
+
|
| 215 |
+
[post-train] loss=0.2989 acc=99.20%
|
| 216 |
+
emb : *0.1567* Δ=-0.0017
|
| 217 |
+
Q_full : *0.1315* Δ=-0.0035
|
| 218 |
+
K_full : 0.1298 Δ=-0.0016
|
| 219 |
+
V_full : *0.1379* Δ=+0.0151
|
| 220 |
+
Q_h0 : 0.1006 Δ=+0.0009
|
| 221 |
+
K_h0 : 0.1167 Δ=-0.0006
|
| 222 |
+
V_h0 : *0.1766* Δ=+0.0320
|
| 223 |
+
Q_h1 : 0.1134 Δ=+0.0041
|
| 224 |
+
K_h1 : 0.1107 Δ=+0.0031
|
| 225 |
+
V_h1 : 0.0841 Δ=-0.0396
|
| 226 |
+
Q_h2 : 0.0947 Δ=-0.0149
|
| 227 |
+
K_h2 : 0.1163 Δ=-0.0000
|
| 228 |
+
V_h2 : 0.1073 Δ=+0.0152
|
| 229 |
+
Q_h3 : 0.1204 Δ=-0.0102
|
| 230 |
+
K_h3 : *0.1500* Δ=+0.0213
|
| 231 |
+
V_h3 : *0.1345* Δ=+0.0301
|
| 232 |
+
out_proj : *0.1462* Δ=+0.0199
|
| 233 |
+
cls_head : 0.0860 Δ=+0.0173
|
| 234 |
+
act_Q_h0 : 0.4820 Δ=+0.0298
|
| 235 |
+
act_K_h0 : 0.7037 Δ=+0.2600
|
| 236 |
+
act_V_h0 : 0.5344 Δ=-0.0138
|
| 237 |
+
act_post_h0 : 0.4416 Δ=-0.0291
|
| 238 |
+
act_Q_h1 : 0.4690 Δ=-0.0262
|
| 239 |
+
act_K_h1 : 0.5742 Δ=+0.0727
|
| 240 |
+
act_V_h1 : 0.5218 Δ=+0.1395
|
| 241 |
+
act_post_h1 : 0.4225 Δ=-0.0761
|
| 242 |
+
act_Q_h2 : 0.4463 Δ=-0.0075
|
| 243 |
+
act_K_h2 : 0.5095 Δ=+0.0613
|
| 244 |
+
act_V_h2 : 0.5113 Δ=+0.0215
|
| 245 |
+
act_post_h2 : 0.4471 Δ=-0.0499
|
| 246 |
+
act_Q_h3 : 0.4629 Δ=-0.0266
|
| 247 |
+
act_K_h3 : 0.5813 Δ=+0.1218
|
| 248 |
+
act_V_h3 : 0.6012 Δ=+0.1740
|
| 249 |
+
act_post_h3 : 0.4323 Δ=-0.0510
|
| 250 |
+
act_emb : *0.2520* Δ=+0.0326
|
| 251 |
+
act_post_full : 0.3062 Δ=-0.0432
|
| 252 |
+
|
| 253 |
+
======================================================================
|
| 254 |
+
D=64 heads=8 head_dim=8
|
| 255 |
+
======================================================================
|
| 256 |
+
|
| 257 |
+
[pre-train]
|
| 258 |
+
emb : *0.1585*
|
| 259 |
+
Q_full : *0.1350*
|
| 260 |
+
K_full : *0.1313*
|
| 261 |
+
V_full : 0.1228
|
| 262 |
+
Q_h0 : 0.0694
|
| 263 |
+
K_h0 : 0.0934
|
| 264 |
+
V_h0 : 0.1261
|
| 265 |
+
Q_h1 : 0.0619
|
| 266 |
+
K_h1 : 0.1018
|
| 267 |
+
V_h1 : 0.0939
|
| 268 |
+
Q_h2 : 0.1182
|
| 269 |
+
K_h2 : 0.0980
|
| 270 |
+
V_h2 : *0.1443*
|
| 271 |
+
Q_h3 : 0.0854
|
| 272 |
+
K_h3 : 0.0866
|
| 273 |
+
V_h3 : 0.0803
|
| 274 |
+
Q_h4 : 0.0612
|
| 275 |
+
K_h4 : 0.1197
|
| 276 |
+
V_h4 : 0.0771
|
| 277 |
+
Q_h5 : 0.0861
|
| 278 |
+
K_h5 : 0.0687
|
| 279 |
+
V_h5 : 0.0657
|
| 280 |
+
Q_h6 : 0.0997
|
| 281 |
+
K_h6 : 0.1160
|
| 282 |
+
V_h6 : 0.0524
|
| 283 |
+
Q_h7 : 0.0983
|
| 284 |
+
K_h7 : 0.1085
|
| 285 |
+
V_h7 : 0.0883
|
| 286 |
+
out_proj : 0.1126
|
| 287 |
+
cls_head : 0.0735
|
| 288 |
+
act_Q_h0 : 0.6606
|
| 289 |
+
act_K_h0 : 0.6856
|
| 290 |
+
act_V_h0 : 0.6908
|
| 291 |
+
act_post_h0 : 0.6642
|
| 292 |
+
act_Q_h1 : 0.5708
|
| 293 |
+
act_K_h1 : 0.6494
|
| 294 |
+
act_V_h1 : 0.6959
|
| 295 |
+
act_post_h1 : 0.6560
|
| 296 |
+
act_Q_h2 : 0.6801
|
| 297 |
+
act_K_h2 : 0.5906
|
| 298 |
+
act_V_h2 : 0.6551
|
| 299 |
+
act_post_h2 : 0.6293
|
| 300 |
+
act_Q_h3 : 0.7066
|
| 301 |
+
act_K_h3 : 0.6817
|
| 302 |
+
act_V_h3 : 0.5788
|
| 303 |
+
act_post_h3 : 0.6498
|
| 304 |
+
act_Q_h4 : 0.6421
|
| 305 |
+
act_K_h4 : 0.5805
|
| 306 |
+
act_V_h4 : 0.7499
|
| 307 |
+
act_post_h4 : 0.7199
|
| 308 |
+
act_Q_h5 : 0.6613
|
| 309 |
+
act_K_h5 : 0.6280
|
| 310 |
+
act_V_h5 : 0.6979
|
| 311 |
+
act_post_h5 : 0.5790
|
| 312 |
+
act_Q_h6 : 0.7152
|
| 313 |
+
act_K_h6 : 0.6223
|
| 314 |
+
act_V_h6 : 0.5662
|
| 315 |
+
act_post_h6 : 0.7522
|
| 316 |
+
act_Q_h7 : 0.7077
|
| 317 |
+
act_K_h7 : 0.6821
|
| 318 |
+
act_V_h7 : 0.6357
|
| 319 |
+
act_post_h7 : 0.6675
|
| 320 |
+
act_emb : *0.2789*
|
| 321 |
+
act_post_full : 0.3142
|
| 322 |
+
|
| 323 |
+
[epoch 25] loss=1.2544 acc=97.40%
|
| 324 |
+
emb : *0.1584*
|
| 325 |
+
Q_full : 0.1223
|
| 326 |
+
K_full : *0.1324*
|
| 327 |
+
V_full : 0.1223
|
| 328 |
+
Q_h0 : 0.0789
|
| 329 |
+
K_h0 : 0.1101
|
| 330 |
+
V_h0 : 0.1208
|
| 331 |
+
Q_h1 : 0.0569
|
| 332 |
+
K_h1 : 0.1061
|
| 333 |
+
V_h1 : 0.1023
|
| 334 |
+
Q_h2 : 0.1134
|
| 335 |
+
K_h2 : 0.1126
|
| 336 |
+
V_h2 : 0.1255
|
| 337 |
+
Q_h3 : 0.0818
|
| 338 |
+
K_h3 : 0.0666
|
| 339 |
+
V_h3 : 0.0566
|
| 340 |
+
Q_h4 : 0.0882
|
| 341 |
+
K_h4 : 0.1170
|
| 342 |
+
V_h4 : 0.0816
|
| 343 |
+
Q_h5 : 0.0915
|
| 344 |
+
K_h5 : 0.0791
|
| 345 |
+
V_h5 : 0.0437
|
| 346 |
+
Q_h6 : 0.0886
|
| 347 |
+
K_h6 : 0.1118
|
| 348 |
+
V_h6 : 0.0766
|
| 349 |
+
Q_h7 : 0.0995
|
| 350 |
+
K_h7 : 0.0995
|
| 351 |
+
V_h7 : 0.0974
|
| 352 |
+
out_proj : *0.1333*
|
| 353 |
+
cls_head : 0.0870
|
| 354 |
+
act_Q_h0 : 0.6729
|
| 355 |
+
act_K_h0 : 0.7706
|
| 356 |
+
act_V_h0 : 0.6221
|
| 357 |
+
act_post_h0 : 0.6544
|
| 358 |
+
act_Q_h1 : 0.6225
|
| 359 |
+
act_K_h1 : 0.6182
|
| 360 |
+
act_V_h1 : 0.6973
|
| 361 |
+
act_post_h1 : 0.5974
|
| 362 |
+
act_Q_h2 : 0.6636
|
| 363 |
+
act_K_h2 : 0.6741
|
| 364 |
+
act_V_h2 : 0.6922
|
| 365 |
+
act_post_h2 : 0.6200
|
| 366 |
+
act_Q_h3 : 0.6775
|
| 367 |
+
act_K_h3 : 0.5720
|
| 368 |
+
act_V_h3 : 0.6252
|
| 369 |
+
act_post_h3 : 0.5886
|
| 370 |
+
act_Q_h4 : 0.7807
|
| 371 |
+
act_K_h4 : 0.6466
|
| 372 |
+
act_V_h4 : 0.7026
|
| 373 |
+
act_post_h4 : 0.5743
|
| 374 |
+
act_Q_h5 : 0.6768
|
| 375 |
+
act_K_h5 : 0.6645
|
| 376 |
+
act_V_h5 : 0.7087
|
| 377 |
+
act_post_h5 : 0.5914
|
| 378 |
+
act_Q_h6 : 0.7016
|
| 379 |
+
act_K_h6 : 0.6182
|
| 380 |
+
act_V_h6 : 0.6397
|
| 381 |
+
act_post_h6 : 0.6251
|
| 382 |
+
act_Q_h7 : 0.7743
|
| 383 |
+
act_K_h7 : 0.6632
|
| 384 |
+
act_V_h7 : 0.8837
|
| 385 |
+
act_post_h7 : 0.6026
|
| 386 |
+
act_emb : *0.2612*
|
| 387 |
+
act_post_full : *0.2900*
|
| 388 |
+
|
| 389 |
+
[post-train] loss=0.3035 acc=99.60%
|
| 390 |
+
emb : *0.1581* Δ=-0.0003
|
| 391 |
+
Q_full : *0.1435* Δ=+0.0084
|
| 392 |
+
K_full : *0.1356* Δ=+0.0043
|
| 393 |
+
V_full : *0.1340* Δ=+0.0112
|
| 394 |
+
Q_h0 : 0.0821 Δ=+0.0127
|
| 395 |
+
K_h0 : 0.1160 Δ=+0.0227
|
| 396 |
+
V_h0 : *0.1440* Δ=+0.0179
|
| 397 |
+
Q_h1 : 0.0821 Δ=+0.0202
|
| 398 |
+
K_h1 : 0.1012 Δ=-0.0007
|
| 399 |
+
V_h1 : *0.1403* Δ=+0.0464
|
| 400 |
+
Q_h2 : *0.1425* Δ=+0.0243
|
| 401 |
+
K_h2 : 0.1203 Δ=+0.0223
|
| 402 |
+
V_h2 : 0.1127 Δ=-0.0316
|
| 403 |
+
Q_h3 : 0.0924 Δ=+0.0070
|
| 404 |
+
K_h3 : 0.0549 Δ=-0.0318
|
| 405 |
+
V_h3 : 0.0873 Δ=+0.0070
|
| 406 |
+
Q_h4 : 0.1184 Δ=+0.0572
|
| 407 |
+
K_h4 : *0.1373* Δ=+0.0176
|
| 408 |
+
V_h4 : 0.1121 Δ=+0.0350
|
| 409 |
+
Q_h5 : 0.1142 Δ=+0.0280
|
| 410 |
+
K_h5 : 0.0933 Δ=+0.0245
|
| 411 |
+
V_h5 : 0.0526 Δ=-0.0131
|
| 412 |
+
Q_h6 : 0.0734 Δ=-0.0263
|
| 413 |
+
K_h6 : 0.1188 Δ=+0.0028
|
| 414 |
+
V_h6 : 0.1089 Δ=+0.0565
|
| 415 |
+
Q_h7 : 0.1029 Δ=+0.0046
|
| 416 |
+
K_h7 : 0.1149 Δ=+0.0064
|
| 417 |
+
V_h7 : 0.1174 Δ=+0.0291
|
| 418 |
+
out_proj : *0.1499* Δ=+0.0373
|
| 419 |
+
cls_head : 0.0871 Δ=+0.0136
|
| 420 |
+
act_Q_h0 : 0.6476 Δ=-0.0129
|
| 421 |
+
act_K_h0 : 0.7738 Δ=+0.0882
|
| 422 |
+
act_V_h0 : 0.6295 Δ=-0.0613
|
| 423 |
+
act_post_h0 : 0.5934 Δ=-0.0708
|
| 424 |
+
act_Q_h1 : 0.6173 Δ=+0.0465
|
| 425 |
+
act_K_h1 : 0.7647 Δ=+0.1153
|
| 426 |
+
act_V_h1 : 0.7091 Δ=+0.0132
|
| 427 |
+
act_post_h1 : 0.6693 Δ=+0.0133
|
| 428 |
+
act_Q_h2 : 0.6804 Δ=+0.0003
|
| 429 |
+
act_K_h2 : 0.6667 Δ=+0.0761
|
| 430 |
+
act_V_h2 : 0.7038 Δ=+0.0487
|
| 431 |
+
act_post_h2 : 0.5798 Δ=-0.0495
|
| 432 |
+
act_Q_h3 : 0.7128 Δ=+0.0062
|
| 433 |
+
act_K_h3 : 0.6496 Δ=-0.0321
|
| 434 |
+
act_V_h3 : 0.7360 Δ=+0.1572
|
| 435 |
+
act_post_h3 : 0.6423 Δ=-0.0075
|
| 436 |
+
act_Q_h4 : 0.7083 Δ=+0.0663
|
| 437 |
+
act_K_h4 : 0.7348 Δ=+0.1543
|
| 438 |
+
act_V_h4 : 0.7565 Δ=+0.0066
|
| 439 |
+
act_post_h4 : 0.5993 Δ=-0.1206
|
| 440 |
+
act_Q_h5 : 0.7219 Δ=+0.0606
|
| 441 |
+
act_K_h5 : 0.6785 Δ=+0.0505
|
| 442 |
+
act_V_h5 : 0.6663 Δ=-0.0317
|
| 443 |
+
act_post_h5 : 0.6103 Δ=+0.0313
|
| 444 |
+
act_Q_h6 : 0.6525 Δ=-0.0627
|
| 445 |
+
act_K_h6 : 0.6987 Δ=+0.0764
|
| 446 |
+
act_V_h6 : 0.6632 Δ=+0.0970
|
| 447 |
+
act_post_h6 : 0.6484 Δ=-0.1038
|
| 448 |
+
act_Q_h7 : 0.6907 Δ=-0.0170
|
| 449 |
+
act_K_h7 : 0.8215 Δ=+0.1394
|
| 450 |
+
act_V_h7 : 0.7420 Δ=+0.1063
|
| 451 |
+
act_post_h7 : 0.6306 Δ=-0.0369
|
| 452 |
+
act_emb : *0.2874* Δ=+0.0085
|
| 453 |
+
act_post_full : 0.3153 Δ=+0.0011
|
| 454 |
+
|
| 455 |
+
======================================================================
|
| 456 |
+
D=128 heads=1 head_dim=128
|
| 457 |
+
======================================================================
|
| 458 |
+
|
| 459 |
+
[pre-train]
|
| 460 |
+
emb : 0.1091
|
| 461 |
+
Q_full : 0.0876
|
| 462 |
+
K_full : 0.0966
|
| 463 |
+
V_full : 0.0937
|
| 464 |
+
Q_h0 : 0.0932
|
| 465 |
+
K_h0 : 0.0978
|
| 466 |
+
V_h0 : 0.1038
|
| 467 |
+
out_proj : 0.0984
|
| 468 |
+
cls_head : 0.0760
|
| 469 |
+
act_Q_h0 : *0.2689*
|
| 470 |
+
act_K_h0 : *0.2891*
|
| 471 |
+
act_V_h0 : *0.2793*
|
| 472 |
+
act_post_h0 : 0.3261
|
| 473 |
+
act_emb : *0.1850*
|
| 474 |
+
act_post_full : 0.3085
|
| 475 |
+
|
| 476 |
+
[epoch 25] loss=0.2378 acc=99.60%
|
| 477 |
+
emb : 0.1124
|
| 478 |
+
Q_full : 0.0901
|
| 479 |
+
K_full : 0.1035
|
| 480 |
+
V_full : 0.0905
|
| 481 |
+
Q_h0 : 0.1002
|
| 482 |
+
K_h0 : 0.1038
|
| 483 |
+
V_h0 : 0.0941
|
| 484 |
+
out_proj : 0.1027
|
| 485 |
+
cls_head : 0.0717
|
| 486 |
+
act_Q_h0 : *0.2432*
|
| 487 |
+
act_K_h0 : 0.3902
|
| 488 |
+
act_V_h0 : 0.3450
|
| 489 |
+
act_post_h0 : *0.2937*
|
| 490 |
+
act_emb : *0.2174*
|
| 491 |
+
act_post_full : 0.3116
|
| 492 |
+
|
| 493 |
+
[post-train] loss=0.0300 acc=100.00%
|
| 494 |
+
emb : 0.1082 Δ=-0.0009
|
| 495 |
+
Q_full : 0.1000 Δ=+0.0124
|
| 496 |
+
K_full : 0.1068 Δ=+0.0102
|
| 497 |
+
V_full : 0.1045 Δ=+0.0108
|
| 498 |
+
Q_h0 : 0.1058 Δ=+0.0125
|
| 499 |
+
K_h0 : 0.1046 Δ=+0.0067
|
| 500 |
+
V_h0 : 0.1070 Δ=+0.0032
|
| 501 |
+
out_proj : 0.1207 Δ=+0.0223
|
| 502 |
+
cls_head : 0.0567 Δ=-0.0193
|
| 503 |
+
act_Q_h0 : 0.3359 Δ=+0.0670
|
| 504 |
+
act_K_h0 : 0.3848 Δ=+0.0958
|
| 505 |
+
act_V_h0 : 0.4073 Δ=+0.1280
|
| 506 |
+
act_post_h0 : 0.3733 Δ=+0.0472
|
| 507 |
+
act_emb : *0.1913* Δ=+0.0063
|
| 508 |
+
act_post_full : 0.3908 Δ=+0.0823
|
| 509 |
+
|
| 510 |
+
======================================================================
|
| 511 |
+
D=128 heads=2 head_dim=64
|
| 512 |
+
======================================================================
|
| 513 |
+
|
| 514 |
+
[pre-train]
|
| 515 |
+
emb : 0.1091
|
| 516 |
+
Q_full : 0.0876
|
| 517 |
+
K_full : 0.0966
|
| 518 |
+
V_full : 0.0937
|
| 519 |
+
Q_h0 : 0.0838
|
| 520 |
+
K_h0 : 0.0998
|
| 521 |
+
V_h0 : 0.0970
|
| 522 |
+
Q_h1 : 0.0834
|
| 523 |
+
K_h1 : 0.0872
|
| 524 |
+
V_h1 : 0.0889
|
| 525 |
+
out_proj : 0.0980
|
| 526 |
+
cls_head : 0.0792
|
| 527 |
+
act_Q_h0 : 0.3140
|
| 528 |
+
act_K_h0 : *0.2951*
|
| 529 |
+
act_V_h0 : 0.3080
|
| 530 |
+
act_post_h0 : 0.3614
|
| 531 |
+
act_Q_h1 : 0.3303
|
| 532 |
+
act_K_h1 : 0.3157
|
| 533 |
+
act_V_h1 : *0.2825*
|
| 534 |
+
act_post_h1 : 0.3438
|
| 535 |
+
act_emb : *0.2059*
|
| 536 |
+
act_post_full : *0.2874*
|
| 537 |
+
|
| 538 |
+
[epoch 25] loss=0.2221 acc=99.60%
|
| 539 |
+
emb : 0.1133
|
| 540 |
+
Q_full : 0.0921
|
| 541 |
+
K_full : 0.0978
|
| 542 |
+
V_full : 0.1024
|
| 543 |
+
Q_h0 : 0.0965
|
| 544 |
+
K_h0 : 0.1039
|
| 545 |
+
V_h0 : 0.0899
|
| 546 |
+
Q_h1 : 0.0880
|
| 547 |
+
K_h1 : 0.0931
|
| 548 |
+
V_h1 : 0.0916
|
| 549 |
+
out_proj : 0.0980
|
| 550 |
+
cls_head : 0.0657
|
| 551 |
+
act_Q_h0 : *0.2908*
|
| 552 |
+
act_K_h0 : 0.3994
|
| 553 |
+
act_V_h0 : 0.3750
|
| 554 |
+
act_post_h0 : 0.3333
|
| 555 |
+
act_Q_h1 : 0.3113
|
| 556 |
+
act_K_h1 : 0.4043
|
| 557 |
+
act_V_h1 : 0.4421
|
| 558 |
+
act_post_h1 : *0.2752*
|
| 559 |
+
act_emb : *0.2419*
|
| 560 |
+
act_post_full : *0.2997*
|
| 561 |
+
|
| 562 |
+
[post-train] loss=0.0202 acc=100.00%
|
| 563 |
+
emb : 0.1077 Δ=-0.0014
|
| 564 |
+
Q_full : 0.0976 Δ=+0.0100
|
| 565 |
+
K_full : 0.1048 Δ=+0.0082
|
| 566 |
+
V_full : 0.1004 Δ=+0.0067
|
| 567 |
+
Q_h0 : 0.0997 Δ=+0.0159
|
| 568 |
+
K_h0 : 0.1130 Δ=+0.0132
|
| 569 |
+
V_h0 : 0.1001 Δ=+0.0031
|
| 570 |
+
Q_h1 : 0.1045 Δ=+0.0210
|
| 571 |
+
K_h1 : 0.0966 Δ=+0.0094
|
| 572 |
+
V_h1 : 0.1054 Δ=+0.0165
|
| 573 |
+
out_proj : 0.1115 Δ=+0.0135
|
| 574 |
+
cls_head : 0.0546 Δ=-0.0246
|
| 575 |
+
act_Q_h0 : 0.3126 Δ=-0.0014
|
| 576 |
+
act_K_h0 : 0.4346 Δ=+0.1395
|
| 577 |
+
act_V_h0 : 0.4026 Δ=+0.0946
|
| 578 |
+
act_post_h0 : 0.3841 Δ=+0.0226
|
| 579 |
+
act_Q_h1 : 0.3240 Δ=-0.0063
|
| 580 |
+
act_K_h1 : 0.4820 Δ=+0.1663
|
| 581 |
+
act_V_h1 : 0.5009 Δ=+0.2184
|
| 582 |
+
act_post_h1 : 0.3587 Δ=+0.0148
|
| 583 |
+
act_emb : *0.2411* Δ=+0.0352
|
| 584 |
+
act_post_full : 0.3226 Δ=+0.0352
|
| 585 |
+
|
| 586 |
+
======================================================================
|
| 587 |
+
D=128 heads=4 head_dim=32
|
| 588 |
+
======================================================================
|
| 589 |
+
|
| 590 |
+
[pre-train]
|
| 591 |
+
emb : 0.1091
|
| 592 |
+
Q_full : 0.0876
|
| 593 |
+
K_full : 0.0966
|
| 594 |
+
V_full : 0.0937
|
| 595 |
+
Q_h0 : 0.0744
|
| 596 |
+
K_h0 : 0.0905
|
| 597 |
+
V_h0 : 0.0977
|
| 598 |
+
Q_h1 : 0.0835
|
| 599 |
+
K_h1 : 0.1030
|
| 600 |
+
V_h1 : 0.0860
|
| 601 |
+
Q_h2 : 0.0694
|
| 602 |
+
K_h2 : 0.0854
|
| 603 |
+
V_h2 : 0.0860
|
| 604 |
+
Q_h3 : 0.0899
|
| 605 |
+
K_h3 : 0.0745
|
| 606 |
+
V_h3 : 0.0898
|
| 607 |
+
out_proj : 0.0956
|
| 608 |
+
cls_head : 0.0776
|
| 609 |
+
act_Q_h0 : 0.3303
|
| 610 |
+
act_K_h0 : 0.3618
|
| 611 |
+
act_V_h0 : 0.3532
|
| 612 |
+
act_post_h0 : 0.4270
|
| 613 |
+
act_Q_h1 : 0.3898
|
| 614 |
+
act_K_h1 : 0.3775
|
| 615 |
+
act_V_h1 : 0.3613
|
| 616 |
+
act_post_h1 : 0.3604
|
| 617 |
+
act_Q_h2 : 0.3821
|
| 618 |
+
act_K_h2 : *0.2836*
|
| 619 |
+
act_V_h2 : 0.4199
|
| 620 |
+
act_post_h2 : 0.4018
|
| 621 |
+
act_Q_h3 : 0.3522
|
| 622 |
+
act_K_h3 : 0.3507
|
| 623 |
+
act_V_h3 : 0.3194
|
| 624 |
+
act_post_h3 : 0.3778
|
| 625 |
+
act_emb : *0.2689*
|
| 626 |
+
act_post_full : 0.3073
|
| 627 |
+
|
| 628 |
+
[epoch 25] loss=0.2307 acc=99.80%
|
| 629 |
+
emb : 0.1132
|
| 630 |
+
Q_full : 0.1062
|
| 631 |
+
K_full : 0.1066
|
| 632 |
+
V_full : 0.1078
|
| 633 |
+
Q_h0 : 0.0894
|
| 634 |
+
K_h0 : 0.1000
|
| 635 |
+
V_h0 : 0.0994
|
| 636 |
+
Q_h1 : 0.0937
|
| 637 |
+
K_h1 : 0.1177
|
| 638 |
+
V_h1 : 0.0893
|
| 639 |
+
Q_h2 : 0.0884
|
| 640 |
+
K_h2 : 0.0969
|
| 641 |
+
V_h2 : 0.0897
|
| 642 |
+
Q_h3 : 0.0934
|
| 643 |
+
K_h3 : 0.0869
|
| 644 |
+
V_h3 : 0.0934
|
| 645 |
+
out_proj : 0.1087
|
| 646 |
+
cls_head : 0.0670
|
| 647 |
+
act_Q_h0 : 0.3745
|
| 648 |
+
act_K_h0 : 0.4438
|
| 649 |
+
act_V_h0 : 0.4299
|
| 650 |
+
act_post_h0 : 0.3798
|
| 651 |
+
act_Q_h1 : 0.3539
|
| 652 |
+
act_K_h1 : 0.5599
|
| 653 |
+
act_V_h1 : 0.3734
|
| 654 |
+
act_post_h1 : 0.3800
|
| 655 |
+
act_Q_h2 : 0.3524
|
| 656 |
+
act_K_h2 : 0.4769
|
| 657 |
+
act_V_h2 : 0.4943
|
| 658 |
+
act_post_h2 : 0.3111
|
| 659 |
+
act_Q_h3 : 0.3788
|
| 660 |
+
act_K_h3 : 0.4899
|
| 661 |
+
act_V_h3 : 0.4093
|
| 662 |
+
act_post_h3 : 0.3438
|
| 663 |
+
act_emb : *0.2119*
|
| 664 |
+
act_post_full : *0.2761*
|
| 665 |
+
|
| 666 |
+
[post-train] loss=0.0187 acc=100.00%
|
| 667 |
+
emb : 0.1128 Δ=+0.0037
|
| 668 |
+
Q_full : 0.1005 Δ=+0.0128
|
| 669 |
+
K_full : 0.1003 Δ=+0.0037
|
| 670 |
+
V_full : 0.1121 Δ=+0.0183
|
| 671 |
+
Q_h0 : 0.0903 Δ=+0.0159
|
| 672 |
+
K_h0 : 0.1076 Δ=+0.0171
|
| 673 |
+
V_h0 : 0.1061 Δ=+0.0084
|
| 674 |
+
Q_h1 : 0.0978 Δ=+0.0143
|
| 675 |
+
K_h1 : 0.1219 Δ=+0.0189
|
| 676 |
+
V_h1 : 0.0991 Δ=+0.0131
|
| 677 |
+
Q_h2 : 0.0963 Δ=+0.0269
|
| 678 |
+
K_h2 : 0.0912 Δ=+0.0058
|
| 679 |
+
V_h2 : 0.0912 Δ=+0.0052
|
| 680 |
+
Q_h3 : 0.0912 Δ=+0.0013
|
| 681 |
+
K_h3 : 0.0962 Δ=+0.0217
|
| 682 |
+
V_h3 : 0.0965 Δ=+0.0067
|
| 683 |
+
out_proj : 0.1104 Δ=+0.0148
|
| 684 |
+
cls_head : 0.0563 Δ=-0.0212
|
| 685 |
+
act_Q_h0 : 0.3274 Δ=-0.0029
|
| 686 |
+
act_K_h0 : 0.5584 Δ=+0.1966
|
| 687 |
+
act_V_h0 : 0.4308 Δ=+0.0776
|
| 688 |
+
act_post_h0 : 0.4036 Δ=-0.0234
|
| 689 |
+
act_Q_h1 : 0.4097 Δ=+0.0199
|
| 690 |
+
act_K_h1 : 0.5285 Δ=+0.1510
|
| 691 |
+
act_V_h1 : 0.4525 Δ=+0.0912
|
| 692 |
+
act_post_h1 : 0.3964 Δ=+0.0361
|
| 693 |
+
act_Q_h2 : 0.3788 Δ=-0.0033
|
| 694 |
+
act_K_h2 : 0.5728 Δ=+0.2892
|
| 695 |
+
act_V_h2 : 0.5443 Δ=+0.1244
|
| 696 |
+
act_post_h2 : 0.3957 Δ=-0.0062
|
| 697 |
+
act_Q_h3 : 0.4042 Δ=+0.0521
|
| 698 |
+
act_K_h3 : 0.5933 Δ=+0.2426
|
| 699 |
+
act_V_h3 : 0.4571 Δ=+0.1376
|
| 700 |
+
act_post_h3 : 0.4140 Δ=+0.0362
|
| 701 |
+
act_emb : *0.2548* Δ=-0.0141
|
| 702 |
+
act_post_full : 0.3589 Δ=+0.0516
|
| 703 |
+
|
| 704 |
+
======================================================================
|
| 705 |
+
D=128 heads=8 head_dim=16
|
| 706 |
+
======================================================================
|
| 707 |
+
|
| 708 |
+
[pre-train]
|
| 709 |
+
emb : 0.1091
|
| 710 |
+
Q_full : 0.0876
|
| 711 |
+
K_full : 0.0966
|
| 712 |
+
V_full : 0.0937
|
| 713 |
+
Q_h0 : 0.0671
|
| 714 |
+
K_h0 : 0.1003
|
| 715 |
+
V_h0 : 0.0769
|
| 716 |
+
Q_h1 : 0.0779
|
| 717 |
+
K_h1 : 0.0858
|
| 718 |
+
V_h1 : 0.0801
|
| 719 |
+
Q_h2 : 0.0705
|
| 720 |
+
K_h2 : 0.1047
|
| 721 |
+
V_h2 : 0.0881
|
| 722 |
+
Q_h3 : 0.0715
|
| 723 |
+
K_h3 : 0.1038
|
| 724 |
+
V_h3 : 0.1002
|
| 725 |
+
Q_h4 : 0.0631
|
| 726 |
+
K_h4 : 0.0919
|
| 727 |
+
V_h4 : 0.0663
|
| 728 |
+
Q_h5 : 0.0736
|
| 729 |
+
K_h5 : 0.0816
|
| 730 |
+
V_h5 : 0.0745
|
| 731 |
+
Q_h6 : 0.0786
|
| 732 |
+
K_h6 : 0.0691
|
| 733 |
+
V_h6 : 0.0865
|
| 734 |
+
Q_h7 : 0.0798
|
| 735 |
+
K_h7 : 0.0788
|
| 736 |
+
V_h7 : 0.0832
|
| 737 |
+
out_proj : 0.0878
|
| 738 |
+
cls_head : 0.0816
|
| 739 |
+
act_Q_h0 : 0.4150
|
| 740 |
+
act_K_h0 : 0.4813
|
| 741 |
+
act_V_h0 : 0.5137
|
| 742 |
+
act_post_h0 : 0.5246
|
| 743 |
+
act_Q_h1 : 0.4347
|
| 744 |
+
act_K_h1 : 0.4596
|
| 745 |
+
act_V_h1 : 0.4736
|
| 746 |
+
act_post_h1 : 0.4873
|
| 747 |
+
act_Q_h2 : 0.4487
|
| 748 |
+
act_K_h2 : 0.4629
|
| 749 |
+
act_V_h2 : 0.4670
|
| 750 |
+
act_post_h2 : 0.5029
|
| 751 |
+
act_Q_h3 : 0.4772
|
| 752 |
+
act_K_h3 : 0.4524
|
| 753 |
+
act_V_h3 : 0.4470
|
| 754 |
+
act_post_h3 : 0.5202
|
| 755 |
+
act_Q_h4 : 0.4767
|
| 756 |
+
act_K_h4 : 0.4430
|
| 757 |
+
act_V_h4 : 0.5214
|
| 758 |
+
act_post_h4 : 0.4237
|
| 759 |
+
act_Q_h5 : 0.4635
|
| 760 |
+
act_K_h5 : 0.4204
|
| 761 |
+
act_V_h5 : 0.5157
|
| 762 |
+
act_post_h5 : 0.4395
|
| 763 |
+
act_Q_h6 : 0.4825
|
| 764 |
+
act_K_h6 : 0.4668
|
| 765 |
+
act_V_h6 : 0.4357
|
| 766 |
+
act_post_h6 : 0.4687
|
| 767 |
+
act_Q_h7 : 0.4366
|
| 768 |
+
act_K_h7 : 0.4388
|
| 769 |
+
act_V_h7 : 0.4442
|
| 770 |
+
act_post_h7 : 0.4719
|
| 771 |
+
act_emb : *0.2489*
|
| 772 |
+
act_post_full : 0.3338
|
| 773 |
+
|
| 774 |
+
[epoch 25] loss=0.2240 acc=99.80%
|
| 775 |
+
emb : 0.1052
|
| 776 |
+
Q_full : 0.0964
|
| 777 |
+
K_full : 0.1000
|
| 778 |
+
V_full : 0.1058
|
| 779 |
+
Q_h0 : 0.0648
|
| 780 |
+
K_h0 : 0.1210
|
| 781 |
+
V_h0 : 0.0831
|
| 782 |
+
Q_h1 : 0.0784
|
| 783 |
+
K_h1 : 0.1015
|
| 784 |
+
V_h1 : 0.0904
|
| 785 |
+
Q_h2 : 0.0818
|
| 786 |
+
K_h2 : 0.0970
|
| 787 |
+
V_h2 : 0.0892
|
| 788 |
+
Q_h3 : 0.0879
|
| 789 |
+
K_h3 : 0.0958
|
| 790 |
+
V_h3 : 0.1011
|
| 791 |
+
Q_h4 : 0.0844
|
| 792 |
+
K_h4 : 0.1019
|
| 793 |
+
V_h4 : 0.0739
|
| 794 |
+
Q_h5 : 0.1001
|
| 795 |
+
K_h5 : 0.0791
|
| 796 |
+
V_h5 : 0.0836
|
| 797 |
+
Q_h6 : 0.0651
|
| 798 |
+
K_h6 : 0.0732
|
| 799 |
+
V_h6 : 0.0937
|
| 800 |
+
Q_h7 : 0.0924
|
| 801 |
+
K_h7 : 0.0984
|
| 802 |
+
V_h7 : 0.1031
|
| 803 |
+
out_proj : 0.0961
|
| 804 |
+
cls_head : 0.0698
|
| 805 |
+
act_Q_h0 : 0.4686
|
| 806 |
+
act_K_h0 : 0.5549
|
| 807 |
+
act_V_h0 : 0.5257
|
| 808 |
+
act_post_h0 : 0.4421
|
| 809 |
+
act_Q_h1 : 0.4416
|
| 810 |
+
act_K_h1 : 0.5945
|
| 811 |
+
act_V_h1 : 0.5244
|
| 812 |
+
act_post_h1 : 0.4517
|
| 813 |
+
act_Q_h2 : 0.4509
|
| 814 |
+
act_K_h2 : 0.6063
|
| 815 |
+
act_V_h2 : 0.4540
|
| 816 |
+
act_post_h2 : 0.4697
|
| 817 |
+
act_Q_h3 : 0.5350
|
| 818 |
+
act_K_h3 : 0.5189
|
| 819 |
+
act_V_h3 : 0.5054
|
| 820 |
+
act_post_h3 : 0.4798
|
| 821 |
+
act_Q_h4 : 0.4223
|
| 822 |
+
act_K_h4 : 0.6416
|
| 823 |
+
act_V_h4 : 0.5742
|
| 824 |
+
act_post_h4 : 0.4320
|
| 825 |
+
act_Q_h5 : 0.4605
|
| 826 |
+
act_K_h5 : 0.5650
|
| 827 |
+
act_V_h5 : 0.5400
|
| 828 |
+
act_post_h5 : 0.4393
|
| 829 |
+
act_Q_h6 : 0.4649
|
| 830 |
+
act_K_h6 : 0.5872
|
| 831 |
+
act_V_h6 : 0.4636
|
| 832 |
+
act_post_h6 : 0.4213
|
| 833 |
+
act_Q_h7 : 0.4619
|
| 834 |
+
act_K_h7 : 0.5432
|
| 835 |
+
act_V_h7 : 0.5168
|
| 836 |
+
act_post_h7 : 0.4866
|
| 837 |
+
act_emb : *0.2009*
|
| 838 |
+
act_post_full : *0.2886*
|
| 839 |
+
|
| 840 |
+
[post-train] loss=0.0188 acc=100.00%
|
| 841 |
+
emb : 0.1060 Δ=-0.0031
|
| 842 |
+
Q_full : 0.0964 Δ=+0.0088
|
| 843 |
+
K_full : 0.0999 Δ=+0.0033
|
| 844 |
+
V_full : 0.1030 Δ=+0.0092
|
| 845 |
+
Q_h0 : 0.0753 Δ=+0.0082
|
| 846 |
+
K_h0 : 0.1260 Δ=+0.0257
|
| 847 |
+
V_h0 : 0.0884 Δ=+0.0116
|
| 848 |
+
Q_h1 : 0.0814 Δ=+0.0035
|
| 849 |
+
K_h1 : 0.1120 Δ=+0.0262
|
| 850 |
+
V_h1 : 0.1015 Δ=+0.0214
|
| 851 |
+
Q_h2 : 0.0873 Δ=+0.0168
|
| 852 |
+
K_h2 : 0.0853 Δ=-0.0194
|
| 853 |
+
V_h2 : 0.0897 Δ=+0.0016
|
| 854 |
+
Q_h3 : 0.1029 Δ=+0.0314
|
| 855 |
+
K_h3 : 0.1043 Δ=+0.0005
|
| 856 |
+
V_h3 : 0.1026 Δ=+0.0023
|
| 857 |
+
Q_h4 : 0.0951 Δ=+0.0320
|
| 858 |
+
K_h4 : 0.1049 Δ=+0.0129
|
| 859 |
+
V_h4 : 0.0763 Δ=+0.0100
|
| 860 |
+
Q_h5 : 0.1152 Δ=+0.0415
|
| 861 |
+
K_h5 : 0.0879 Δ=+0.0063
|
| 862 |
+
V_h5 : 0.0949 Δ=+0.0205
|
| 863 |
+
Q_h6 : 0.0790 Δ=+0.0003
|
| 864 |
+
K_h6 : 0.0789 Δ=+0.0098
|
| 865 |
+
V_h6 : 0.0932 Δ=+0.0067
|
| 866 |
+
Q_h7 : 0.0970 Δ=+0.0172
|
| 867 |
+
K_h7 : 0.0939 Δ=+0.0151
|
| 868 |
+
V_h7 : 0.1120 Δ=+0.0288
|
| 869 |
+
out_proj : 0.1142 Δ=+0.0264
|
| 870 |
+
cls_head : 0.0556 Δ=-0.0260
|
| 871 |
+
act_Q_h0 : 0.4538 Δ=+0.0388
|
| 872 |
+
act_K_h0 : 0.6259 Δ=+0.1445
|
| 873 |
+
act_V_h0 : 0.5796 Δ=+0.0659
|
| 874 |
+
act_post_h0 : 0.5135 Δ=-0.0111
|
| 875 |
+
act_Q_h1 : 0.4504 Δ=+0.0157
|
| 876 |
+
act_K_h1 : 0.6123 Δ=+0.1527
|
| 877 |
+
act_V_h1 : 0.5364 Δ=+0.0628
|
| 878 |
+
act_post_h1 : 0.4148 Δ=-0.0724
|
| 879 |
+
act_Q_h2 : 0.4330 Δ=-0.0158
|
| 880 |
+
act_K_h2 : 0.6079 Δ=+0.1450
|
| 881 |
+
act_V_h2 : 0.5365 Δ=+0.0695
|
| 882 |
+
act_post_h2 : 0.4718 Δ=-0.0311
|
| 883 |
+
act_Q_h3 : 0.4942 Δ=+0.0170
|
| 884 |
+
act_K_h3 : 0.7313 Δ=+0.2788
|
| 885 |
+
act_V_h3 : 0.4719 Δ=+0.0249
|
| 886 |
+
act_post_h3 : 0.4444 Δ=-0.0758
|
| 887 |
+
act_Q_h4 : 0.5071 Δ=+0.0304
|
| 888 |
+
act_K_h4 : 0.7608 Δ=+0.3178
|
| 889 |
+
act_V_h4 : 0.5955 Δ=+0.0741
|
| 890 |
+
act_post_h4 : 0.4462 Δ=+0.0225
|
| 891 |
+
act_Q_h5 : 0.4960 Δ=+0.0324
|
| 892 |
+
act_K_h5 : 0.6651 Δ=+0.2447
|
| 893 |
+
act_V_h5 : 0.5983 Δ=+0.0826
|
| 894 |
+
act_post_h5 : 0.4369 Δ=-0.0026
|
| 895 |
+
act_Q_h6 : 0.5170 Δ=+0.0345
|
| 896 |
+
act_K_h6 : 0.6135 Δ=+0.1467
|
| 897 |
+
act_V_h6 : 0.6220 Δ=+0.1863
|
| 898 |
+
act_post_h6 : 0.4884 Δ=+0.0197
|
| 899 |
+
act_Q_h7 : 0.4773 Δ=+0.0407
|
| 900 |
+
act_K_h7 : 0.6479 Δ=+0.2091
|
| 901 |
+
act_V_h7 : 0.5699 Δ=+0.1257
|
| 902 |
+
act_post_h7 : 0.4634 Δ=-0.0085
|
| 903 |
+
act_emb : *0.2147* Δ=-0.0342
|
| 904 |
+
act_post_full : 0.3396 Δ=+0.0058
|
| 905 |
+
|
| 906 |
+
======================================================================
|
| 907 |
+
D=256 heads=1 head_dim=256
|
| 908 |
+
======================================================================
|
| 909 |
+
|
| 910 |
+
[pre-train]
|
| 911 |
+
emb : 0.0779
|
| 912 |
+
Q_full : 0.0682
|
| 913 |
+
K_full : 0.0633
|
| 914 |
+
V_full : 0.0665
|
| 915 |
+
Q_h0 : 0.0654
|
| 916 |
+
K_h0 : 0.0627
|
| 917 |
+
V_h0 : 0.0663
|
| 918 |
+
out_proj : 0.0616
|
| 919 |
+
cls_head : 0.0492
|
| 920 |
+
act_Q_h0 : *0.2094*
|
| 921 |
+
act_K_h0 : *0.2437*
|
| 922 |
+
act_V_h0 : *0.2355*
|
| 923 |
+
act_post_h0 : *0.2695*
|
| 924 |
+
act_emb : *0.2625*
|
| 925 |
+
act_post_full : 0.3118
|
| 926 |
+
|
| 927 |
+
[epoch 25] loss=0.0099 acc=100.00%
|
| 928 |
+
emb : 0.0821
|
| 929 |
+
Q_full : 0.0664
|
| 930 |
+
K_full : 0.0681
|
| 931 |
+
V_full : 0.0728
|
| 932 |
+
Q_h0 : 0.0728
|
| 933 |
+
K_h0 : 0.0676
|
| 934 |
+
V_h0 : 0.0716
|
| 935 |
+
out_proj : 0.0762
|
| 936 |
+
cls_head : 0.0455
|
| 937 |
+
act_Q_h0 : *0.2878*
|
| 938 |
+
act_K_h0 : 0.4158
|
| 939 |
+
act_V_h0 : 0.3795
|
| 940 |
+
act_post_h0 : 0.3578
|
| 941 |
+
act_emb : *0.1934*
|
| 942 |
+
act_post_full : 0.3588
|
| 943 |
+
|
| 944 |
+
[post-train] loss=0.0016 acc=100.00%
|
| 945 |
+
emb : 0.0819 Δ=+0.0040
|
| 946 |
+
Q_full : 0.0716 Δ=+0.0033
|
| 947 |
+
K_full : 0.0785 Δ=+0.0151
|
| 948 |
+
V_full : 0.0790 Δ=+0.0125
|
| 949 |
+
Q_h0 : 0.0720 Δ=+0.0066
|
| 950 |
+
K_h0 : 0.0728 Δ=+0.0102
|
| 951 |
+
V_h0 : 0.0783 Δ=+0.0120
|
| 952 |
+
out_proj : 0.0775 Δ=+0.0159
|
| 953 |
+
cls_head : 0.0486 Δ=-0.0006
|
| 954 |
+
act_Q_h0 : *0.2487* Δ=+0.0393
|
| 955 |
+
act_K_h0 : 0.4001 Δ=+0.1563
|
| 956 |
+
act_V_h0 : 0.4007 Δ=+0.1652
|
| 957 |
+
act_post_h0 : 0.4446 Δ=+0.1751
|
| 958 |
+
act_emb : *0.2063* Δ=-0.0563
|
| 959 |
+
act_post_full : 0.4458 Δ=+0.1339
|
| 960 |
+
|
| 961 |
+
======================================================================
|
| 962 |
+
D=256 heads=2 head_dim=128
|
| 963 |
+
======================================================================
|
| 964 |
+
|
| 965 |
+
[pre-train]
|
| 966 |
+
emb : 0.0779
|
| 967 |
+
Q_full : 0.0682
|
| 968 |
+
K_full : 0.0633
|
| 969 |
+
V_full : 0.0665
|
| 970 |
+
Q_h0 : 0.0632
|
| 971 |
+
K_h0 : 0.0698
|
| 972 |
+
V_h0 : 0.0634
|
| 973 |
+
Q_h1 : 0.0612
|
| 974 |
+
K_h1 : 0.0588
|
| 975 |
+
V_h1 : 0.0607
|
| 976 |
+
out_proj : 0.0637
|
| 977 |
+
cls_head : 0.0487
|
| 978 |
+
act_Q_h0 : *0.2391*
|
| 979 |
+
act_K_h0 : *0.2624*
|
| 980 |
+
act_V_h0 : *0.2857*
|
| 981 |
+
act_post_h0 : 0.3254
|
| 982 |
+
act_Q_h1 : *0.2333*
|
| 983 |
+
act_K_h1 : *0.2212*
|
| 984 |
+
act_V_h1 : *0.2717*
|
| 985 |
+
act_post_h1 : *0.2777*
|
| 986 |
+
act_emb : *0.2220*
|
| 987 |
+
act_post_full : *0.2732*
|
| 988 |
+
|
| 989 |
+
[epoch 25] loss=0.0079 acc=100.00%
|
| 990 |
+
emb : 0.0767
|
| 991 |
+
Q_full : 0.0683
|
| 992 |
+
K_full : 0.0663
|
| 993 |
+
V_full : 0.0692
|
| 994 |
+
Q_h0 : 0.0661
|
| 995 |
+
K_h0 : 0.0700
|
| 996 |
+
V_h0 : 0.0681
|
| 997 |
+
Q_h1 : 0.0724
|
| 998 |
+
K_h1 : 0.0710
|
| 999 |
+
V_h1 : 0.0707
|
| 1000 |
+
out_proj : 0.0748
|
| 1001 |
+
cls_head : 0.0475
|
| 1002 |
+
act_Q_h0 : 0.3045
|
| 1003 |
+
act_K_h0 : 0.4500
|
| 1004 |
+
act_V_h0 : 0.4044
|
| 1005 |
+
act_post_h0 : 0.3539
|
| 1006 |
+
act_Q_h1 : *0.2901*
|
| 1007 |
+
act_K_h1 : 0.4922
|
| 1008 |
+
act_V_h1 : 0.3736
|
| 1009 |
+
act_post_h1 : 0.3901
|
| 1010 |
+
act_emb : *0.2386*
|
| 1011 |
+
act_post_full : 0.3775
|
| 1012 |
+
|
| 1013 |
+
[post-train] loss=0.0014 acc=100.00%
|
| 1014 |
+
emb : 0.0804 Δ=+0.0025
|
| 1015 |
+
Q_full : 0.0749 Δ=+0.0066
|
| 1016 |
+
K_full : 0.0742 Δ=+0.0109
|
| 1017 |
+
V_full : 0.0755 Δ=+0.0090
|
| 1018 |
+
Q_h0 : 0.0721 Δ=+0.0089
|
| 1019 |
+
K_h0 : 0.0799 Δ=+0.0101
|
| 1020 |
+
V_h0 : 0.0743 Δ=+0.0109
|
| 1021 |
+
Q_h1 : 0.0743 Δ=+0.0130
|
| 1022 |
+
K_h1 : 0.0784 Δ=+0.0196
|
| 1023 |
+
V_h1 : 0.0776 Δ=+0.0169
|
| 1024 |
+
out_proj : 0.0859 Δ=+0.0222
|
| 1025 |
+
cls_head : 0.0480 Δ=-0.0007
|
| 1026 |
+
act_Q_h0 : *0.2627* Δ=+0.0237
|
| 1027 |
+
act_K_h0 : 0.4986 Δ=+0.2362
|
| 1028 |
+
act_V_h0 : 0.4207 Δ=+0.1350
|
| 1029 |
+
act_post_h0 : 0.4344 Δ=+0.1090
|
| 1030 |
+
act_Q_h1 : *0.2992* Δ=+0.0659
|
| 1031 |
+
act_K_h1 : 0.4905 Δ=+0.2693
|
| 1032 |
+
act_V_h1 : 0.4743 Δ=+0.2025
|
| 1033 |
+
act_post_h1 : 0.4108 Δ=+0.1331
|
| 1034 |
+
act_emb : *0.2547* Δ=+0.0327
|
| 1035 |
+
act_post_full : 0.3783 Δ=+0.1052
|
| 1036 |
+
|
| 1037 |
+
======================================================================
|
| 1038 |
+
D=256 heads=4 head_dim=64
|
| 1039 |
+
======================================================================
|
| 1040 |
+
|
| 1041 |
+
[pre-train]
|
| 1042 |
+
emb : 0.0779
|
| 1043 |
+
Q_full : 0.0682
|
| 1044 |
+
K_full : 0.0633
|
| 1045 |
+
V_full : 0.0665
|
| 1046 |
+
Q_h0 : 0.0598
|
| 1047 |
+
K_h0 : 0.0591
|
| 1048 |
+
V_h0 : 0.0664
|
| 1049 |
+
Q_h1 : 0.0568
|
| 1050 |
+
K_h1 : 0.0580
|
| 1051 |
+
V_h1 : 0.0588
|
| 1052 |
+
Q_h2 : 0.0655
|
| 1053 |
+
K_h2 : 0.0629
|
| 1054 |
+
V_h2 : 0.0583
|
| 1055 |
+
Q_h3 : 0.0582
|
| 1056 |
+
K_h3 : 0.0558
|
| 1057 |
+
V_h3 : 0.0657
|
| 1058 |
+
out_proj : 0.0656
|
| 1059 |
+
cls_head : 0.0482
|
| 1060 |
+
act_Q_h0 : *0.2595*
|
| 1061 |
+
act_K_h0 : 0.3228
|
| 1062 |
+
act_V_h0 : *0.2912*
|
| 1063 |
+
act_post_h0 : *0.2840*
|
| 1064 |
+
act_Q_h1 : *0.2824*
|
| 1065 |
+
act_K_h1 : 0.3022
|
| 1066 |
+
act_V_h1 : *0.2934*
|
| 1067 |
+
act_post_h1 : 0.3500
|
| 1068 |
+
act_Q_h2 : *0.2818*
|
| 1069 |
+
act_K_h2 : *0.2558*
|
| 1070 |
+
act_V_h2 : 0.3219
|
| 1071 |
+
act_post_h2 : 0.3397
|
| 1072 |
+
act_Q_h3 : 0.3140
|
| 1073 |
+
act_K_h3 : 0.3174
|
| 1074 |
+
act_V_h3 : *0.2822*
|
| 1075 |
+
act_post_h3 : 0.3068
|
| 1076 |
+
act_emb : *0.2217*
|
| 1077 |
+
act_post_full : *0.2799*
|
| 1078 |
+
|
| 1079 |
+
[epoch 25] loss=0.0074 acc=100.00%
|
| 1080 |
+
emb : 0.0756
|
| 1081 |
+
Q_full : 0.0707
|
| 1082 |
+
K_full : 0.0690
|
| 1083 |
+
V_full : 0.0749
|
| 1084 |
+
Q_h0 : 0.0658
|
| 1085 |
+
K_h0 : 0.0739
|
| 1086 |
+
V_h0 : 0.0764
|
| 1087 |
+
Q_h1 : 0.0661
|
| 1088 |
+
K_h1 : 0.0672
|
| 1089 |
+
V_h1 : 0.0695
|
| 1090 |
+
Q_h2 : 0.0743
|
| 1091 |
+
K_h2 : 0.0738
|
| 1092 |
+
V_h2 : 0.0738
|
| 1093 |
+
Q_h3 : 0.0724
|
| 1094 |
+
K_h3 : 0.0596
|
| 1095 |
+
V_h3 : 0.0745
|
| 1096 |
+
out_proj : 0.0732
|
| 1097 |
+
cls_head : 0.0456
|
| 1098 |
+
act_Q_h0 : 0.3024
|
| 1099 |
+
act_K_h0 : 0.5494
|
| 1100 |
+
act_V_h0 : 0.4222
|
| 1101 |
+
act_post_h0 : 0.3549
|
| 1102 |
+
act_Q_h1 : *0.2892*
|
| 1103 |
+
act_K_h1 : 0.4484
|
| 1104 |
+
act_V_h1 : 0.4565
|
| 1105 |
+
act_post_h1 : 0.4029
|
| 1106 |
+
act_Q_h2 : *0.2954*
|
| 1107 |
+
act_K_h2 : 0.5268
|
| 1108 |
+
act_V_h2 : 0.4681
|
| 1109 |
+
act_post_h2 : 0.4058
|
| 1110 |
+
act_Q_h3 : *0.2740*
|
| 1111 |
+
act_K_h3 : 0.5939
|
| 1112 |
+
act_V_h3 : 0.4381
|
| 1113 |
+
act_post_h3 : 0.4051
|
| 1114 |
+
act_emb : *0.1759*
|
| 1115 |
+
act_post_full : 0.3800
|
| 1116 |
+
|
| 1117 |
+
[post-train] loss=0.0013 acc=100.00%
|
| 1118 |
+
emb : 0.0859 Δ=+0.0080
|
| 1119 |
+
Q_full : 0.0735 Δ=+0.0053
|
| 1120 |
+
K_full : 0.0713 Δ=+0.0079
|
| 1121 |
+
V_full : 0.0773 Δ=+0.0108
|
| 1122 |
+
Q_h0 : 0.0678 Δ=+0.0079
|
| 1123 |
+
K_h0 : 0.0733 Δ=+0.0141
|
| 1124 |
+
V_h0 : 0.0779 Δ=+0.0115
|
| 1125 |
+
Q_h1 : 0.0735 Δ=+0.0168
|
| 1126 |
+
K_h1 : 0.0774 Δ=+0.0194
|
| 1127 |
+
V_h1 : 0.0732 Δ=+0.0144
|
| 1128 |
+
Q_h2 : 0.0739 Δ=+0.0084
|
| 1129 |
+
K_h2 : 0.0844 Δ=+0.0215
|
| 1130 |
+
V_h2 : 0.0843 Δ=+0.0260
|
| 1131 |
+
Q_h3 : 0.0789 Δ=+0.0207
|
| 1132 |
+
K_h3 : 0.0666 Δ=+0.0108
|
| 1133 |
+
V_h3 : 0.0767 Δ=+0.0111
|
| 1134 |
+
out_proj : 0.0836 Δ=+0.0180
|
| 1135 |
+
cls_head : 0.0488 Δ=+0.0006
|
| 1136 |
+
act_Q_h0 : 0.3580 Δ=+0.0986
|
| 1137 |
+
act_K_h0 : 0.5870 Δ=+0.2643
|
| 1138 |
+
act_V_h0 : 0.4358 Δ=+0.1446
|
| 1139 |
+
act_post_h0 : 0.4285 Δ=+0.1446
|
| 1140 |
+
act_Q_h1 : *0.2921* Δ=+0.0097
|
| 1141 |
+
act_K_h1 : 0.5681 Δ=+0.2658
|
| 1142 |
+
act_V_h1 : 0.4575 Δ=+0.1641
|
| 1143 |
+
act_post_h1 : 0.4672 Δ=+0.1172
|
| 1144 |
+
act_Q_h2 : 0.3152 Δ=+0.0334
|
| 1145 |
+
act_K_h2 : 0.5638 Δ=+0.3080
|
| 1146 |
+
act_V_h2 : 0.4609 Δ=+0.1390
|
| 1147 |
+
act_post_h2 : 0.4649 Δ=+0.1252
|
| 1148 |
+
act_Q_h3 : 0.3445 Δ=+0.0306
|
| 1149 |
+
act_K_h3 : 0.5860 Δ=+0.2685
|
| 1150 |
+
act_V_h3 : 0.4535 Δ=+0.1713
|
| 1151 |
+
act_post_h3 : 0.4180 Δ=+0.1112
|
| 1152 |
+
act_emb : *0.2088* Δ=-0.0129
|
| 1153 |
+
act_post_full : 0.4113 Δ=+0.1314
|
| 1154 |
+
|
| 1155 |
+
======================================================================
|
| 1156 |
+
D=256 heads=8 head_dim=32
|
| 1157 |
+
======================================================================
|
| 1158 |
+
|
| 1159 |
+
[pre-train]
|
| 1160 |
+
emb : 0.0779
|
| 1161 |
+
Q_full : 0.0682
|
| 1162 |
+
K_full : 0.0633
|
| 1163 |
+
V_full : 0.0665
|
| 1164 |
+
Q_h0 : 0.0603
|
| 1165 |
+
K_h0 : 0.0687
|
| 1166 |
+
V_h0 : 0.0600
|
| 1167 |
+
Q_h1 : 0.0554
|
| 1168 |
+
K_h1 : 0.0675
|
| 1169 |
+
V_h1 : 0.0623
|
| 1170 |
+
Q_h2 : 0.0658
|
| 1171 |
+
K_h2 : 0.0613
|
| 1172 |
+
V_h2 : 0.0529
|
| 1173 |
+
Q_h3 : 0.0479
|
| 1174 |
+
K_h3 : 0.0583
|
| 1175 |
+
V_h3 : 0.0627
|
| 1176 |
+
Q_h4 : 0.0559
|
| 1177 |
+
K_h4 : 0.0609
|
| 1178 |
+
V_h4 : 0.0645
|
| 1179 |
+
Q_h5 : 0.0694
|
| 1180 |
+
K_h5 : 0.0643
|
| 1181 |
+
V_h5 : 0.0547
|
| 1182 |
+
Q_h6 : 0.0669
|
| 1183 |
+
K_h6 : 0.0576
|
| 1184 |
+
V_h6 : 0.0527
|
| 1185 |
+
Q_h7 : 0.0648
|
| 1186 |
+
K_h7 : 0.0610
|
| 1187 |
+
V_h7 : 0.0662
|
| 1188 |
+
out_proj : 0.0628
|
| 1189 |
+
cls_head : 0.0472
|
| 1190 |
+
act_Q_h0 : 0.3938
|
| 1191 |
+
act_K_h0 : 0.3478
|
| 1192 |
+
act_V_h0 : *0.2891*
|
| 1193 |
+
act_post_h0 : 0.4032
|
| 1194 |
+
act_Q_h1 : 0.3324
|
| 1195 |
+
act_K_h1 : 0.3333
|
| 1196 |
+
act_V_h1 : 0.3331
|
| 1197 |
+
act_post_h1 : 0.3885
|
| 1198 |
+
act_Q_h2 : 0.3593
|
| 1199 |
+
act_K_h2 : *0.2980*
|
| 1200 |
+
act_V_h2 : *0.2990*
|
| 1201 |
+
act_post_h2 : 0.3510
|
| 1202 |
+
act_Q_h3 : 0.3019
|
| 1203 |
+
act_K_h3 : 0.3741
|
| 1204 |
+
act_V_h3 : 0.3351
|
| 1205 |
+
act_post_h3 : 0.4049
|
| 1206 |
+
act_Q_h4 : 0.3307
|
| 1207 |
+
act_K_h4 : 0.3595
|
| 1208 |
+
act_V_h4 : 0.3515
|
| 1209 |
+
act_post_h4 : 0.4145
|
| 1210 |
+
act_Q_h5 : 0.3736
|
| 1211 |
+
act_K_h5 : 0.3552
|
| 1212 |
+
act_V_h5 : 0.3845
|
| 1213 |
+
act_post_h5 : 0.3626
|
| 1214 |
+
act_Q_h6 : 0.3142
|
| 1215 |
+
act_K_h6 : *0.2974*
|
| 1216 |
+
act_V_h6 : 0.3430
|
| 1217 |
+
act_post_h6 : 0.3921
|
| 1218 |
+
act_Q_h7 : *0.2749*
|
| 1219 |
+
act_K_h7 : 0.3761
|
| 1220 |
+
act_V_h7 : 0.3521
|
| 1221 |
+
act_post_h7 : 0.3747
|
| 1222 |
+
act_emb : *0.2389*
|
| 1223 |
+
act_post_full : 0.3172
|
| 1224 |
+
|
| 1225 |
+
[epoch 25] loss=0.0076 acc=100.00%
|
| 1226 |
+
emb : 0.0870
|
| 1227 |
+
Q_full : 0.0710
|
| 1228 |
+
K_full : 0.0713
|
| 1229 |
+
V_full : 0.0733
|
| 1230 |
+
Q_h0 : 0.0603
|
| 1231 |
+
K_h0 : 0.0704
|
| 1232 |
+
V_h0 : 0.0705
|
| 1233 |
+
Q_h1 : 0.0652
|
| 1234 |
+
K_h1 : 0.0863
|
| 1235 |
+
V_h1 : 0.0722
|
| 1236 |
+
Q_h2 : 0.0668
|
| 1237 |
+
K_h2 : 0.0676
|
| 1238 |
+
V_h2 : 0.0678
|
| 1239 |
+
Q_h3 : 0.0557
|
| 1240 |
+
K_h3 : 0.0716
|
| 1241 |
+
V_h3 : 0.0721
|
| 1242 |
+
Q_h4 : 0.0678
|
| 1243 |
+
K_h4 : 0.0681
|
| 1244 |
+
V_h4 : 0.0854
|
| 1245 |
+
Q_h5 : 0.0660
|
| 1246 |
+
K_h5 : 0.0748
|
| 1247 |
+
V_h5 : 0.0727
|
| 1248 |
+
Q_h6 : 0.0757
|
| 1249 |
+
K_h6 : 0.0613
|
| 1250 |
+
V_h6 : 0.0645
|
| 1251 |
+
Q_h7 : 0.0725
|
| 1252 |
+
K_h7 : 0.0721
|
| 1253 |
+
V_h7 : 0.0784
|
| 1254 |
+
out_proj : 0.0805
|
| 1255 |
+
cls_head : 0.0470
|
| 1256 |
+
act_Q_h0 : 0.3898
|
| 1257 |
+
act_K_h0 : 0.6606
|
| 1258 |
+
act_V_h0 : 0.4376
|
| 1259 |
+
act_post_h0 : 0.3801
|
| 1260 |
+
act_Q_h1 : 0.3430
|
| 1261 |
+
act_K_h1 : 0.5881
|
| 1262 |
+
act_V_h1 : 0.5103
|
| 1263 |
+
act_post_h1 : 0.4429
|
| 1264 |
+
act_Q_h2 : 0.3749
|
| 1265 |
+
act_K_h2 : 0.6678
|
| 1266 |
+
act_V_h2 : 0.5165
|
| 1267 |
+
act_post_h2 : 0.4137
|
| 1268 |
+
act_Q_h3 : 0.3357
|
| 1269 |
+
act_K_h3 : 0.5528
|
| 1270 |
+
act_V_h3 : 0.4197
|
| 1271 |
+
act_post_h3 : 0.4182
|
| 1272 |
+
act_Q_h4 : 0.3857
|
| 1273 |
+
act_K_h4 : 0.6191
|
| 1274 |
+
act_V_h4 : 0.5490
|
| 1275 |
+
act_post_h4 : 0.4189
|
| 1276 |
+
act_Q_h5 : 0.3940
|
| 1277 |
+
act_K_h5 : 0.6308
|
| 1278 |
+
act_V_h5 : 0.4108
|
| 1279 |
+
act_post_h5 : 0.4423
|
| 1280 |
+
act_Q_h6 : 0.3946
|
| 1281 |
+
act_K_h6 : 0.5846
|
| 1282 |
+
act_V_h6 : 0.4491
|
| 1283 |
+
act_post_h6 : 0.3772
|
| 1284 |
+
act_Q_h7 : 0.4124
|
| 1285 |
+
act_K_h7 : 0.6258
|
| 1286 |
+
act_V_h7 : 0.4564
|
| 1287 |
+
act_post_h7 : 0.3559
|
| 1288 |
+
act_emb : *0.2063*
|
| 1289 |
+
act_post_full : 0.3397
|
| 1290 |
+
|
| 1291 |
+
[post-train] loss=0.0013 acc=100.00%
|
| 1292 |
+
emb : 0.0809 Δ=+0.0030
|
| 1293 |
+
Q_full : 0.0691 Δ=+0.0008
|
| 1294 |
+
K_full : 0.0710 Δ=+0.0077
|
| 1295 |
+
V_full : 0.0852 Δ=+0.0187
|
| 1296 |
+
Q_h0 : 0.0667 Δ=+0.0064
|
| 1297 |
+
K_h0 : 0.0788 Δ=+0.0100
|
| 1298 |
+
V_h0 : 0.0728 Δ=+0.0128
|
| 1299 |
+
Q_h1 : 0.0674 Δ=+0.0119
|
| 1300 |
+
K_h1 : 0.0844 Δ=+0.0169
|
| 1301 |
+
V_h1 : 0.0812 Δ=+0.0189
|
| 1302 |
+
Q_h2 : 0.0730 Δ=+0.0073
|
| 1303 |
+
K_h2 : 0.0781 Δ=+0.0168
|
| 1304 |
+
V_h2 : 0.0715 Δ=+0.0186
|
| 1305 |
+
Q_h3 : 0.0678 Δ=+0.0199
|
| 1306 |
+
K_h3 : 0.0716 Δ=+0.0133
|
| 1307 |
+
V_h3 : 0.0764 Δ=+0.0137
|
| 1308 |
+
Q_h4 : 0.0713 Δ=+0.0154
|
| 1309 |
+
K_h4 : 0.0740 Δ=+0.0131
|
| 1310 |
+
V_h4 : 0.0909 Δ=+0.0264
|
| 1311 |
+
Q_h5 : 0.0712 Δ=+0.0018
|
| 1312 |
+
K_h5 : 0.0801 Δ=+0.0158
|
| 1313 |
+
V_h5 : 0.0697 Δ=+0.0151
|
| 1314 |
+
Q_h6 : 0.0804 Δ=+0.0135
|
| 1315 |
+
K_h6 : 0.0658 Δ=+0.0082
|
| 1316 |
+
V_h6 : 0.0647 Δ=+0.0120
|
| 1317 |
+
Q_h7 : 0.0754 Δ=+0.0106
|
| 1318 |
+
K_h7 : 0.0699 Δ=+0.0090
|
| 1319 |
+
V_h7 : 0.0800 Δ=+0.0138
|
| 1320 |
+
out_proj : 0.0846 Δ=+0.0218
|
| 1321 |
+
cls_head : 0.0466 Δ=-0.0007
|
| 1322 |
+
act_Q_h0 : 0.4054 Δ=+0.0116
|
| 1323 |
+
act_K_h0 : 0.6899 Δ=+0.3421
|
| 1324 |
+
act_V_h0 : 0.5066 Δ=+0.2175
|
| 1325 |
+
act_post_h0 : 0.4566 Δ=+0.0535
|
| 1326 |
+
act_Q_h1 : 0.3412 Δ=+0.0087
|
| 1327 |
+
act_K_h1 : 0.7009 Δ=+0.3676
|
| 1328 |
+
act_V_h1 : 0.6057 Δ=+0.2726
|
| 1329 |
+
act_post_h1 : 0.4266 Δ=+0.0381
|
| 1330 |
+
act_Q_h2 : 0.3666 Δ=+0.0073
|
| 1331 |
+
act_K_h2 : 0.6611 Δ=+0.3632
|
| 1332 |
+
act_V_h2 : 0.5442 Δ=+0.2452
|
| 1333 |
+
act_post_h2 : 0.4957 Δ=+0.1447
|
| 1334 |
+
act_Q_h3 : 0.3971 Δ=+0.0951
|
| 1335 |
+
act_K_h3 : 0.6011 Δ=+0.2270
|
| 1336 |
+
act_V_h3 : 0.4752 Δ=+0.1402
|
| 1337 |
+
act_post_h3 : 0.4357 Δ=+0.0308
|
| 1338 |
+
act_Q_h4 : 0.3645 Δ=+0.0338
|
| 1339 |
+
act_K_h4 : 0.5758 Δ=+0.2163
|
| 1340 |
+
act_V_h4 : 0.5611 Δ=+0.2096
|
| 1341 |
+
act_post_h4 : 0.4185 Δ=+0.0040
|
| 1342 |
+
act_Q_h5 : 0.3831 Δ=+0.0095
|
| 1343 |
+
act_K_h5 : 0.6076 Δ=+0.2524
|
| 1344 |
+
act_V_h5 : 0.4840 Δ=+0.0995
|
| 1345 |
+
act_post_h5 : 0.4087 Δ=+0.0461
|
| 1346 |
+
act_Q_h6 : 0.3817 Δ=+0.0675
|
| 1347 |
+
act_K_h6 : 0.5986 Δ=+0.3012
|
| 1348 |
+
act_V_h6 : 0.5178 Δ=+0.1749
|
| 1349 |
+
act_post_h6 : 0.4694 Δ=+0.0773
|
| 1350 |
+
act_Q_h7 : 0.3795 Δ=+0.1046
|
| 1351 |
+
act_K_h7 : 0.5845 Δ=+0.2084
|
| 1352 |
+
act_V_h7 : 0.4897 Δ=+0.1376
|
| 1353 |
+
act_post_h7 : 0.4559 Δ=+0.0811
|
| 1354 |
+
act_emb : *0.2226* Δ=-0.0164
|
| 1355 |
+
act_post_full : 0.4059 Δ=+0.0887
|
| 1356 |
+
|
| 1357 |
+
|
| 1358 |
+
======================================================================
|
| 1359 |
+
SUMMARY: Post-training WEIGHT CV by head_dim
|
| 1360 |
+
======================================================================
|
| 1361 |
+
D heads hdim | emb Q_full K_full V_full out | acc
|
| 1362 |
+
--------------------------------------------------------------------------------
|
| 1363 |
+
64 1 64 | *0.1549* *0.1352* *0.1453* 0.1219 *0.1486* | 98.80%
|
| 1364 |
+
64 2 32 | *0.1606* 0.1187 *0.1422* 0.1286 *0.1468* | 99.20%
|
| 1365 |
+
64 4 16 | *0.1567* *0.1315* 0.1298 *0.1379* *0.1462* | 99.20%
|
| 1366 |
+
64 8 8 | *0.1581* *0.1435* *0.1356* *0.1340* *0.1499* | 99.60%
|
| 1367 |
+
128 1 128 | 0.1082 0.1000 0.1068 0.1045 0.1207 | 100.00%
|
| 1368 |
+
128 2 64 | 0.1077 0.0976 0.1048 0.1004 0.1115 | 100.00%
|
| 1369 |
+
128 4 32 | 0.1128 0.1005 0.1003 0.1121 0.1104 | 100.00%
|
| 1370 |
+
128 8 16 | 0.1060 0.0964 0.0999 0.1030 0.1142 | 100.00%
|
| 1371 |
+
256 1 256 | 0.0819 0.0716 0.0785 0.0790 0.0775 | 100.00%
|
| 1372 |
+
256 2 128 | 0.0804 0.0749 0.0742 0.0755 0.0859 | 100.00%
|
| 1373 |
+
256 4 64 | 0.0859 0.0735 0.0713 0.0773 0.0836 | 100.00%
|
| 1374 |
+
256 8 32 | 0.0809 0.0691 0.0710 0.0852 0.0846 | 100.00%
|
| 1375 |
+
|
| 1376 |
+
|
| 1377 |
+
======================================================================
|
| 1378 |
+
SUMMARY: Post-training ACTIVATION CV by head_dim
|
| 1379 |
+
(These measure the space where attention actually operates)
|
| 1380 |
+
======================================================================
|
| 1381 |
+
D heads hdim | act_emb aQ_h0 aK_h0 aV_h0 aPost0 act_full | acc
|
| 1382 |
+
------------------------------------------------------------------------------------------
|
| 1383 |
+
64 1 64 | *0.2970* 0.3235 0.4174 0.4063 0.3038 0.3234 | 98.80%
|
| 1384 |
+
64 2 32 | *0.2472* 0.3814 0.4524 0.4259 0.3327 *0.2776* | 99.20%
|
| 1385 |
+
64 4 16 | *0.2520* 0.4820 0.7037 0.5344 0.4416 0.3062 | 99.20%
|
| 1386 |
+
64 8 8 | *0.2874* 0.6476 0.7738 0.6295 0.5934 0.3153 | 99.60%
|
| 1387 |
+
128 1 128 | *0.1913* 0.3359 0.3848 0.4073 0.3733 0.3908 | 100.00%
|
| 1388 |
+
128 2 64 | *0.2411* 0.3126 0.4346 0.4026 0.3841 0.3226 | 100.00%
|
| 1389 |
+
128 4 32 | *0.2548* 0.3274 0.5584 0.4308 0.4036 0.3589 | 100.00%
|
| 1390 |
+
128 8 16 | *0.2147* 0.4538 0.6259 0.5796 0.5135 0.3396 | 100.00%
|
| 1391 |
+
256 1 256 | *0.2063* *0.2487* 0.4001 0.4007 0.4446 0.4458 | 100.00%
|
| 1392 |
+
256 2 128 | *0.2547* *0.2627* 0.4986 0.4207 0.4344 0.3783 | 100.00%
|
| 1393 |
+
256 4 64 | *0.2088* 0.3580 0.5870 0.4358 0.4285 0.4113 | 100.00%
|
| 1394 |
+
256 8 32 | *0.2226* 0.4054 0.6899 0.5066 0.4566 0.4059 | 100.00%
|
| 1395 |
+
|
| 1396 |
+
|
| 1397 |
+
======================================================================
|
| 1398 |
+
SUMMARY: ACTIVATION CV movement (post - pre)
|
| 1399 |
+
======================================================================
|
| 1400 |
+
D heads hdim | act_emb aQ_h0 aK_h0 aV_h0 aPost0 act_full
|
| 1401 |
+
--------------------------------------------------------------------------------
|
| 1402 |
+
64 1 64 | +0.0149 -0.0201 +0.0810 +0.0836 +0.0199 +0.0026
|
| 1403 |
+
64 2 32 | +0.0158 +0.0153 +0.1042 +0.0627 -0.0669 -0.0006
|
| 1404 |
+
64 4 16 | +0.0326 +0.0298 +0.2600 -0.0138 -0.0291 -0.0432
|
| 1405 |
+
64 8 8 | +0.0085 -0.0129 +0.0882 -0.0613 -0.0708 +0.0011
|
| 1406 |
+
128 1 128 | +0.0063 +0.0670 +0.0958 +0.1280 +0.0472 +0.0823
|
| 1407 |
+
128 2 64 | +0.0352 -0.0014 +0.1395 +0.0946 +0.0226 +0.0352
|
| 1408 |
+
128 4 32 | -0.0141 -0.0029 +0.1966 +0.0776 -0.0234 +0.0516
|
| 1409 |
+
128 8 16 | -0.0342 +0.0388 +0.1445 +0.0659 -0.0111 +0.0058
|
| 1410 |
+
256 1 256 | -0.0563 +0.0393 +0.1563 +0.1652 +0.1751 +0.1339
|
| 1411 |
+
256 2 128 | +0.0327 +0.0237 +0.2362 +0.1350 +0.1090 +0.1052
|
| 1412 |
+
256 4 64 | -0.0129 +0.0986 +0.2643 +0.1446 +0.1446 +0.1314
|
| 1413 |
+
256 8 32 | -0.0164 +0.0116 +0.3421 +0.2175 +0.0535 +0.0887
|