CODEWITHSUNDEEP
openai-apipre-trained-modelgpt-3fine-tuning
epahomov
epahomov

Reputation: 651

Fine-tuning of OpeanAI model with unsupervised set, not supervised

I want GPT-3 model to know everything about my domain area, for example my inbox. I want to be able to ask it questions like "Have I even had a Silicon Valley Bank account?" and get correct response. I've familiarized myself with fine-tuning mechanism in official OpenAI docs and it's not exactly what I'm looking for. I want to just dump all my emails on the model and ask it: "Learn!". However fine-tuning require supervised style learning with prompts and reponses, which I do not have. Example in the notebooks for doc suggests that you can use "Davinci-instruct to ask a few questions based on a Wikipedia section, as well as answer those questions, based on that section", which I guess solves my problem if I apply it to all my emails, but I'd rather not do this step, because I might screw up something. Can I have other options?

I found that Azure Open AI integration allows you to do fine-tuning as well, but it seems to have the same problem.

I might be calling what I want to do is fine-tuning, but in fact I keep pre-training process and just decided to go with fine-tuning because it has documentation and API. On the other hand fine-tuning guaranties that I would get wrong answers, pre-training doesn't, and you dont want to get wrong answer on question "Have I even had a Silicon Valley Bank account?"

Upvotes: 2

Views: 937

Answers (1)

Valentin
Valentin

Reputation: 365

You can generate training set of prompt/completion pairs based on your knowledge base via various techniques such as:

  • summarizations (for example via GPT models);
  • generated Q/A via GPT;
  • cloze-style questions (again you can utilize GPT itself to suggest such questions);

Use augmentations (paraphrasing, word substitutions, changing the order of sentences).

Upvotes: 1

Related Questions