#Fine-tuning to imitate someone

3 messages · Page 1 of 1 (latest)

brazen stag
#

Hey,

I want to use OpenAI's fine-tuning API to be able to make it so that I can imitate somebody, make GPT write like them, have the same opinons as them, etc. I have thousands of tweets that this person has tweeted, along with what they're quoting, etc.

I'm really confused on where to start and I have these questions:

  1. If I were to fine-tune using tweets this person has written, what would I put the prompt as in the data?
  2. After fine tuning, how would I get GPT to create completely random sentences, statements/opinions,
  3. Is it possible I could get GPT to make up conversations between this person and other people? Would it be possible for me to use the API to simulate conversations, or social media interactions?
  4. This is a stupid question and I'm sorry but is there anyway to make the completions api respond with parseable data? Like, could I get it to reply with JSON objects?
wintry bridge
#
  1. I think that openAI API response is native JSON.
#
  1. to train or finetune, you deliver a sequence of tokens from the data and you seek a response, and use the ground truth next tokens as basis for computing LOSS for gradient back propagation? Eventually, you train the model to specialize on the new user data. P.S. OpenAI currently has shortage of GPUs. So investigate if a downloadable GPT2 or similar GPT model can run on your local hardware. If local GPT model works well enough finetuned with your user data then you don't need to use the API.