Alpaca instruction mode findings

[anzz1]

From my testing, it seems that using the ### Instruction: ### Response: isn't strictly necessary.

Anecdotally, the fine-tuning done with the instruction-response model seems to have taught it to be better at answering prompts even when the instruct-response model is not used.

Testing with the chat-with-bob.txt prompt (not using instruct mode), the regular llama model often leads to interactions where Bob answers with "I can do that" or similar, and you need to input "Go on" or something to get the actual answer. With alpaca models I haven't seen such behaviour, and it always directly answers the questions asked.

That being said, it would make logical sense that using the instruct mode should give better output. However, I don't know if that's the case. In any case, the training clearly has an effect outside the instruct-response context too.