Fine-tuningYou fine-tune a 65B base model into an assistant with supervised examples. Doubling the training set from the same source doesn't improve the responses.
Most of the model's knowledge comes from pretraining; fine-tuning mainly teaches format and style.