I have an application that uses multiple different GPT API calls with different system prompts, but I’m wondering does that mean I should train different models for each different system prompt or can i train one model that has a variety of different system prompts?
Training different models for a specialization may be best, but that precludes common uses outside of what can be inferred by interpolating between examples.
It depends on if you application will be switching system prompts - and how effective that switch is.
“You are ChatGPT” as a system message, you get an AI that can write poems or attempt orbital dynamics questions. The switch of behavior is the user input.
If you can diverge your training even further from the chat skill and other subtasks with a unique system message, then better for you. The more unique the system prompt sequence, the more the AI will be put into a “mode” instead of having “AI talks like a pirate” bleed into “AI talks like a robot” — or “AI already thinks it is ChatGPT”.
1 Like
It makes sense then to train different models for different aspects of operation.
A “what temperature does this question need” AI, or a “does this try to jailbreak” AI, will have responses incompatible with “here’s a chatbot that checks our product database first”.