Fixing It in Post: AComparative Study of LLM Post-Training Data Quality and Model Performance