Fine-tuning multiple equally valid answers

API
,
1

There are examples of fine-tuning in the documentation at OpenAI Platform.

What does the JSON format look like for questions where there are multiple equally valid answers?

Then, for the same question

  1. are the possible answers simply listed in a separate record or
  2. are the answers concatenated or
  3. is there a special format for this?
2

Hi,

No special format, you can list the same question with multiple answers i.e. Q:what is the sun A: the sun is a star Q:what is the sun A:the sun is a hot ball of plasma. or you could try giving multiple “assistant” roll replies to one query… not sure how that would work performance wise, but I do not think that would be syntactically incorrect.

2 Likes
3

Those won’t have an equal chance after they are in the model, though, they will be affected by the context.

“My friend was telling me that the sun wasn’t a star, because stars only come out at night. What is the sun?”

“In relation to the four states of matter: what is the sun?”

The fine-tuning file can even have multiples of the same question if you particularly want to train on that behavior, for example to overcome the built-in training of gpt-3.5-turbo to deny and disclaim.

2 Likes