How to reduce rms error in network training (FANN & php)?

Question

Neural networks

I train the network on 20,000 lines.
Next, I give 5000 new ones for verification.
The accuracy is 65-66%.
Tried many times, this result is the best.
I read that the default system error is 0.35.
Found how to get and how to reset it.
Therefore, the question arose, what should I do with this error and at what stage of training? (As I understand it, it needs to be counted, let's say, every 10-20 training iterations, correct me if I'm wrong)
I found the info, how is fann_test_data() , but I don't quite understand what should be the second parameter? What kind of data and how to organize it as a resource?
UPD 1:
I tried to use the functionfann_test() - returns the same as fann_run() .
It is not entirely clear how I can correct the error, or rather the debugging mechanism itself.