Truly Testing LLM Impersonators

Aligned models maintain much of their aligned style even after finetuning. If you finetune a base model on a large portion of a specific person's data, how predictive is it? Perplexity is not a good measure here, as it will mostly be descriptive of style. But I think it's interesting to ask if LLMs really could guess how people would react in novel situations. Too bad we don't have Diary Corpus. Yet.