I am working on a project to mimic a real person’s style of texting using GPT. I did some research on the problem and i concluded that fine-tuning is good for tasks that involve domain-specific knowledge that may not be present in the pre-trained language model and even that fine-tuning can sometimes lower the model’s performance in some cases. and im coming more and more to the conclusion that prompting is the best way to achieve my goal. by prompting I mean feeding the model set of pairs of person’s answers to chats with each request. (data i have is a chat history of person's chat in whatssap, 100-200 pairs of chats)
#Fine-tuning or Prompting? what's the optimal way to mimic a personal style of texting?
abstract gull
livid marten
I am working on a project to mimic a real person’s style of texting using GPT. I...
According to OpenAI's guidelines, prompt engineering is a recommended starting point because it allows for quicker iterations and immediate feedback. For your project, where you aim to mimic a person's texting style with a dataset of 100-200 chat pairs, prompt engineering seems particularly suitable. You can swiftly test and refine prompts to approximate the desired texting style.
However, it's worth noting that fine-tuning, while effective in some scenarios, comes with its own set of drawbacks. It requires a substantial investment of time and resources, including the creation of training datasets and running training jobs. The feedback loop is also much slower compared to prompt engineering, making it less flexible for quick iterations.
In summary, given your specific needs and dataset size, starting off with prompt engineering could offer a more agile approach. Fine-tuning could be considered later if prompt engineering doesn't fully meet your requirements, but be mindful of the time and effort it demands.
summarized by gpt-4 from https://platform.openai.com/docs/guides/fine-tuning
abstract gull
thanks for the response. im kind leaning towards that conclusion too. one thing thats holding me back is that maybe sending 100 pairs of Q&A is not the most cost-effective way to do it. im interested in your take on that ! thanks again
mild quiver
If you do go with fine tuning, you might find this helpful https://www.npmjs.com/package/@kenzic/story
