how to decrease RMS training error of force in lcurve.out?

[Wangmumu0403]

I have tried to increase the NN neuron to [240 240 240] and incrase the start_pref_f to 2000 and limit_pref_f is 1.2 and axis_neoron to 1.6 ; while these operations not work for rmse_f_trn in lcurve.out.
At the same time . I also check my Dataset, I want to ask if the force's fluctuation range will influence the rmse_f_trn?
which part will help for my decrease the rmse_f_trn ?

#########
Here is my rmse_f_trn in lcurve.out.

and .json file
#input.json.docx
and dataset Raw file's force file 's the first column

[Wangmumu0403]

Q3: Does the discrete point will influence the force training?

[njzjz]

I notice your rcut_smth is larger than rcut, which may result in unusual behavior. rcut_smth should be less than rcut.

[Wangmumu0403]

Thanks ! I will test this parameter again .