A development team has a large language model that was initially trained on a vast and diverse collection of text from the internet, enabling it to understand grammar, facts, and reasoning. The team's new goal is to adapt this model to become a specialized and helpful assistant for writing professional business emails. Which of the following data strategies for this second phase of training would be most effective in achieving this specific goal?