1Dataset cleaning,bad rows removed.Dedupe rows and fix bad labels, drop corrupt uploads before training.
2Run comparison,ranked by metric.Compare runs on accuracy and loss, then ship the checkpoint that wins.