Using GPT-3 to generate text according to a pre-defined heuristic


I’m currently working on a building a fine-tuned model on GPT-3 that takes a sentence as input and and generates a new sentence based on a pre-defined ‘heuristic’.

The data was initially in an .xlxs format, and after preprocessing it I’ve created a “prompt” and “completion” column for the fine-tuning later. The “prompt” column essentially consists of the original sentences and the “completion” column consists of the heuristicized sentences. The original sentences have all been selected according to a particular heuristic.

I had a couple of questions regarding this problem:

  1. Is the current approach feasible for solving this problem? If not, then are there any other routes that could potentially work?
  2. I’ve converted this dataframe into a .jsonl file but I’m a bit confused about the next steps. How exactly do you incorporate this into a fine-tuned GPT-3 model?
  3. What GPT-3 model would be the best choice for this problem?
  4. Is there any way to include the heuristic used as a parameter while setting up the fine-tuned model?

Thanks in advance! Let me know if there’s anything I need to clear up.

As the definition of heuristic is behavior mediated by intuition and experience, yes, building up a fine-tuned dataset should work for this. I have many repos demonstrating finetuning as well as videos on YouTube