Hey, guys! I’m new with AI and could really use some guidance. We’ve got a dataset containing 50,000 headlines of significant news, and our goal is to train ChatGPT to recognize them as important. The idea is to provide ChatGPT with a new headline and receive a relevance score from 1 to 10, reflecting its importance based on the understanding developed from the 50,000 headlines. Any suggestions on how to make this happen?
My suggestion would be to create a fine-tuned GPT-3.5-turbo model. In your system message, describe the task and explain the scoring mechanism. Use as training examples the headline as user message and the score as assistant output. Preparing a training data set of 100-200 examples on the basis of which you create the model should give you a solid start.
Take a look at the following OpenAI guidance for further details on fine-tuning: