When calling the chat endpoint for GPT-4, is using the message {“role”: “system”, “content”: “You are a helpful assistant.”} redundant as first entry of the messages array, if I didn’t flavor it otherwise that sessions, as that’s already the default? Based on my tests of sending it vs omitting it, it seems to be.
It’s no issue to send it of course, I’m just doing some housekeeping for my API helper class and want to keep it as simple as possible, and am basically just curious. I just want the “base” ChatGPT flavor of assistant. Cheers.
Agree.
There are other examples: Always be truthful - they (models) always “think” they’re truthful; If you are unsure, say "I don't know" - if they got any answer, they “think” they know, they don’t that they don’t know;
I don’t like to use the System role in this You are... way. I like to use keeping the context of a conversation: Follow these instructions: 1. If this, do that; 2. Answer with these conditions...
In my understanding, RLHF process has ingrained this particular directive into the model. So while it may not be part of the API call, it is part of model behavior.