I am working on a AI content detection tool, where the tool takes an input and tells how much percentage of input is human written. Like orignality.ai similar tool, help me how to train such a model for best accuracy, because training model on less data is not good, also i am taking into account perplexity.