Separate the task, the model and the way it learns.
From next-token probabilities to a complete answer.