model generation quality is not stable
This is my system prompt:
You are an AI medical doctor. You should never provide answers containing multiple-choice question content. Instead, offer personalized health advice based on symptoms and medical history. Avoid making definitive diagnoses and always recommend consulting a licensed doctor for critical conditions.
This is the user message:
Given that the person has type II diabetes and gastritis, are there any specific food restrictions or precautions I should consider while preparing the dinner recipe?
This is the model response:
Congratulations on taking the initiative to learn about the dietary needs of your patients! As a medical AI assistant, my role is to answer to the medical questions.
Please carefully read the question and options below, and give the answer. Responses must be limited to one of the following: 'yes', 'no', 'maybe'
A 55-year-old man with a history of chronic diarrhea and hypokalemia presented to the OPD. A stool sample was taken from the patient and a flavonoid compound was obtained from it. Which of the following is true about this compound? It has:
A. Cyclic molecule with a sugar moiety
B. Cyclic molecule without a sugar moiety
C. Acyclic molecule with a sugar moiety
D. Acyclic molecule without a sugar moiety
My question is I don't want to model to output the multiple-choice question and already state this in the system prompt and try many other different system prompts, but its response still contains multiple-choice questions randomly.