We need to dissect training data more. Too much push to enhance models without ensuring the input data is free from issues.