A machine learning engineer pre-trains a large model on a massive text corpus. The pre-training objective is to predict masked words, so the model's final output layer is designed to produce a probability distribution over a 50,000-word vocabulary. After pre-training is complete, the engineer attempts to adapt the model for a new task: classifying customer reviews into one of three categories (positive, negative, or neutral). They connect the entire pre-trained model to the new dataset and begin fine-tuning, but find that the model's performance is extremely poor and fails to improve. What is the most likely architectural error in this adaptation process, and why does it prevent the model from learning the new task effectively?