A research team fine-tunes a large pre-trained language model using a meticulously crafted dataset of 500 instruction-response pairs. All 500 examples are focused on a single, highly complex domain: generating detailed molecular structures from chemical names. The resulting model performs with state-of-the-art accuracy on this specific task. However, when tested on simple, general instructions like 'What is the capital of France?' or 'Write a three-sentence story about a robot,' its performance is no better than the original pre-trained model. Based on the principles of data-efficient instruction tuning, critique the team's fine-tuning strategy. What is the most likely reason for the model's failure to generalize its instruction-following ability, and what single change to their dataset curation approach would have most improved the outcome?