>>51137847its its *SOMETHING* like this
temp is a value that tries to smooth out the difference between tokens
high temp = more slope-y = less 'ideal' answers brought up, more 'ideal' answers brought down
low temp = more mountain-y = more 'ideal' answers stay up, less 'ideal' answers stay down
top P acts as a cutoff line that actually starts at the bottom with a value of '1'
as the number gets smaller the line goes up
if a token is below the line, the model straight up ignores its existence
if its above, the model will *consider* it as a 'valid' answer but still may not choose it
top K is the chance the model will pick a neighboring token in vector space instead of an 'ideal' answer
like
>what color is the sky"blue" is an 'ideal' answer
but "red" and "orange" are also *valid* but unlikely answers located near where "blue" is located at
higher values = more lul so randum because your sort of but not actually telling it
>i want you to be more creative than correct 