I want to build an application with GPT4 API tailored to my specific applications. This means that I do not want unwanted outputs from it even if it is able to produce them. An example would be Khan Academy’s use case. You would want the chatbot to guide the user to discover the correct answer and hence learn something from the process, as opposed to directly outputting the solution, even the chatbot is able to do so. I am not sure how to do this, since I supposed we cannot fine-tune the model from just an API. One possible way could be to prompt it at the very beginning, but how to make sure people wouldn’t jailbreak it? Is there other more secure methods to do so?
The best way is to definitely prompt it in the system and user messages to tell it exactly what you want it accomplish. Adding a few examples would help as well. While fine tuning is not possible for any model except base davinci right now, you could provide it with some of the necessary information for the task as a part of the context window, which is 8k tokens for GPT
In addition to the previous response: I included several advices in this How to Limit question results to proprietary dataset? - #2 by AgusPG. Hope it helps
Thank you for the information provided! I’ll look into them
Looks very comprehensive! I’ll look into them thanks a lot