Hi Yu,
I set 'wxc_scale = 0.1', and also included secondary structure restraints, reference model.
Using Ramachandran restraints may be a good idea to use at this resolution too.
The 'individual site' still give me higher Rfree and bigger gap between Rfree and Rwork of my dataset (~3.8A).
"Higher" and "bigger" than what? Your statement above does not contain any information for me to comment on.
Is the 'wxc_scale = 0.1' not tight enough?
If the default setting does not work for you, then you have two options: 1) let phenix.refine find optimal values for you automatically, using optimize_wxc=true and optimize_wxu=true or if this doesn't work for you too, then 2) keep trying different values for wxc_scale and wxu_scale until you hit the one that produces the results you like. Pavel.