{“role”: “assistant”, “content”: f"here are the topics that I want you to classify, seperated by ‘,’. here the topics:{str_topics}. it’s important that all the classifications will be from the list i gave you exactlly." },
First - look at the role. The AI appears to be saying to the user “I want you to classify”. The user role tells the AI. Or if system programming of behavior, put it all in a system role.
Outside of a few spelling mistakes reducing comprehension, I would use programming techniques which the AI will understand.
system: from article, find topic. output format: “Topic: best_topic_from_list”
user: choose from the most applicable article topic. topic_list = [“aliens”, “pyramids”, “magic crystals”,…]
user: extract topic: {your_article}
This assistant role you can now toss completely, as it adds no value, and we replaced the others:
{“role”: “assistant”, “content”: f"you are a classification analysis assistant. I have a list of topics that interests me. given a text, I want you to check if the text is talking about any of those topics Whether directly or indirectly, in the spirit of things (but classif by the only topics i will give you) and output the topics the texttalked about."}
If the latest gpt-3.5-turbo fails you badly, try -0301