Hi everyone. Following up on my previous question. So I created an app that creates quiz questions, with 4 multiple choice answers. Users select an answer and move on to the next question. But I’m having two persistent problems: GPT returns very easy questions (even though my prompt instructs GPT to search for a “high difficulty and lesser-known fact” and to cross-check the veracity of the fact before publishing it as a quiz question. Nevertheless, GTP from time to time spits up totally factually wrong questions and answers. For example, it will ask a question as to which actor was the lead in a well-known movie and the 4 possible answers do not name the correct actor. That kind of thing. Also, the questions are super obvious, like what is the capital of Canada. Anyway, just wondering if there is any way to structure the prompt (I’m using the API) to reduce false questions or answers and to increase the level of question difficulty. Any advice is appreciated.
Factual accuracy is still an issue with most of the generative AI, although it’s ameliorating slowly.
You can try passing contextual knowledge using a block of text to a model and then prompting the model to generate questions based only on that context.
That should improve the factual accuracy of your generated questions.
1 Like
Thanks for the reply, but that approach does not make the app feasible. But thanks for the thoughts.