Encountered abnormal ultra-high precision when using dpgen2 sampling

[JiangXiaoMingSan]

I used dpgen2 and DPA-2 for sampling, and used the TiVZrNb dataset with a dpgen sampling accuracy of 95% as input to train the potential function. I sampled the ternary alloy composed of elements in the dataset, and used fixed levels as convergence parameters, "level_f_hi": 0.35,"level_f_lo": At 0.15, the accuracy reaches 100%, and the training steps are 100000. I increased the training steps to 200000 and the accuracy became 97%. This is very confusing for me because I used 1000 steps as input when testing the program, and the accuracy obtained at this time is also 100%. In my understanding, such a small number of steps should result in a very poor model. Why is there such high accuracy? Or should I use adaptive power as the convergence parameter? Do you have any suggestions? Is this normal?
step=100000
stage id_stg. iter. accu. cand. fail. lvl_f_lo lvl_f_hi cvged
Stage 0 --------------------
0 0 0 1.0000 0.0000 0.0000 0.1500 0.3500 True
Stage 0 converged YES reached max numb iterations NO
All stages converged
step=200000
stage id_stg. iter. accu. cand. fail. lvl_f_lo lvl_f_hi cvged
Stage 0 --------------------
0 0 0 0.9760 0.0202 0.0037 0.1500 0.3500 True
Stage 0 converged YES reached max numb iterations NO
All stages converged
I only modified the training step size for these two different processes, and the script is as follows. Thank you for your attention to this issue, it is very important to me.
train.json
template.lammps.json
input.json