Fine-tuning only available for 'base models'?

API
1

Hi

The OpenAI documentation about fine tuning states ‘Fine-tuning is currently only available for the following base models: davinci, curie, babbage, and ada. These are the original models that do not have any instruction following training (like text-davinci-003 does for example).’.

Can anyone help me understand what exactly this means? Specifically, my questions are

  1. Does the last sentence in the quote above means that you cannot ‘prompt’ the fine tuned model like you do with gpt-3.5-turbo?
  2. Following the above, does that mean that the fine tuned model can only do the task you fine tuned it for? I.e., the same as whatever ‘prompt/completion’ pairs you used for fine tuning?
  3. The base models are quite outdated but I may be wrong… however, if that is indeed the case, what would be a use case to use fine tuning other than the more advanced prompting models like gpt-3.5-turbo, gpt 4?

Thanks

2
  1. You can prompt a fine-tuned model exactly as you would any other.
  2. You can ask a fine-tuned model to do anything you want, how well it will do what you ask is an altogether different question.
  3. The point is control over the response. Fine-tuning will have a strong influence on how the model responds to the prompt.
1 Like
3

Thanks for the insights, very helpful!

4

If you want to feed specific knowledge into the model e.g. to search stuff you might want to have a look into embeddings instead.

5

Thank you. I read about it and I think my use case of email generation given some prompts/parameters isn’t a good fit for embedding based approaches. But do correct me if I am wrong.

What I need to do is to describe certain ‘characteristics’ of an email (e.g., emotion, sentiment, story focus, purpose etc) and ask GPT to create an email. We have a lot of such data that can be used as training data for fine tuning, imagine one example being a given email, and multiple labels (over 20) assigned to it. It seems to me that prompting may be the best way to start, but fine tuning may be useful to teach the model what each characteristic means…

Thank you for your insights again.

6

That’s right. To change the tone, etc. you might have to finetune. Maybe even a combination could be useful (using two different models for that would be cheaper i suppose).