A startup is developing a general-purpose AI assistant using a pre-trained large language model. Their goal is to create a single model that can handle a wide variety of user requests, such as drafting emails, summarizing articles, and answering factual questions. They have a limited budget for data collection and model training. They are considering two different strategies for creating their fine-tuning dataset:

• Strategy A: Create a single, unified dataset that mixes together thousands of examples of email drafting, article summarization, and question-answering, with each example framed as a specific instruction (e.g., "Summarize the following text:", "Draft an email to...").

• Strategy B: Create three separate, highly specialized datasets—one exclusively for email drafting, one for summarization, and one for question-answering. They would then train three separate, specialized models.

Given the startup's goal of creating a single, versatile AI assistant with a limited budget, which strategy should they choose? Justify your recommendation by explaining the likely impact of the chosen data strategy on the model's final capabilities.