i cannot understand how to train on custom dataset . and the dataset used to train this is open-sourced ?