Hello, Thank you for this great contribution. Can you please clarify recommendations for training on a custom dataset?