Hi @Rodrigo_M!
Unfortunately I am unable to directly support local model training. However, to get the best results, I suggest having a test folder with, to start, 10% of the images in your complete dataset.
We generally recommend starting with a 70/20/10 split between the train/valid/test splits.