Questions for the combination of deepmd-kit and dpgen.

[composites123]

I currently have around 18,000 frames of data for different atomic systems, which can be used for the initial training of deepmd-kit. How can I proceed to further supplement and expand the dataset using dpgen?
In the param.json file of dpgen, I need to specify the initial structure POSCAR for the system. However, my initial data contains a large variety of different systems. How should I address this step?