##### ###### ##### ### # # ### # # ###### ## ## ## ## ## ## ## # # # # # ## ##### #### ##### # # # # # # # #### ## # ## ## ## ## # # # # # ## ## # ###### ## ### # ### # ######
##### ###### ##### ### # # ### # # ###### ## ## ## ## ## ## ## # # # # # ## ##### #### ##### # # # # # # # #### ## # ## ## ## ## # # # # # ## ## # ###### ## ### # ### # ######
Alpha (α) scales the LoRA contribution:
W_new = W_original + (α/r) × BA
Higher alpha means stronger adaptation. The ratio α/r determines the actual scaling factor.
Common practice: Set α = r (so scaling = ) or α = r (scaling = ). If r = , try α = or α = . The optimal value depends on your task and data.
