How do I overcome model degradation in Generative AI models when training on non-ideal datasets like noisy text data

Question

Can i know How do I overcome model degradation in Generative AI models when training on non-ideal datasets like noisy text data?