Could adding iteration's fp as a new system leads to data leakages?

[dmh1998dmh]

When a iteration finished, labelled data will be trained as a seperated system in the next iteration. DPGEN makes only one set in 02.fp/data.000 folder. Deepmd-kit documents said that deepmd trains mutlisystems one by one and chooses test frames randomly from dataset if there is only one set (this case).
I am wondering whether there is a data leakage in the new system and models be overfitted to the new system.
A learning lcurve is attached.

The learning lcurve is not smooth, and is it possible sudden changes result from the overfit to addition dataset?

[AnguseZhang]

Learning curve not smooth: It's quite common since in a training process, systems are randomly selected and there's no guarantees for the monotonous decrease.
Data leakage : Not fully understand what do you mean. New labels are trained equally with old systems, if you do not use the "init-model" mode. You can use dp test to check your model quality.

[dmh1998dmh]

Thanks for your reply, What i mean in data leakage is that one label is used both in training and testing. One paper (DOI:10.1021/acs.chemrev.0c01111) said it would cause problems.