Fine-tuning - How to deal with bullets in the answer text?

API
,
1

What is the best way to implement enumerations in the content area when fine-tuning?

  1. As a comma-separated list,
  2. preceded by a bullet?
  3. on a separate line?

Is there a “best practice” for this?

2

ChatGPT and thus the underlying gpt-3.5-turbo are pretrained on producing markdown format.

Markdown is relatively simple, a minus sign (-) and a space, and you’ll begin a bullet unordered list. A number, a period, and a space, and that instead is a numbered list. It is the same format this forum uses, which you can test in the reply box.

The basic list structure can also be displayed as plain text without need for a UI interpreter (while bold text with two asterisks surrounding it must be rendered).

You can certainly demonstrate other outputs in your fine-tune examples, even training the AI to write HTML, but that’s a larger hurdle.

1 Like
3

Personally I find it performs better if:

  • numbered list in instructions,
  • bullet list in answers when items are long or have their own syntax
  • comma separated if list and items are short in answers
  • comma separated list if answer to summarizing to a list of subjects/keywords
  • bullet list of summarizing to longer items

But then again, it’s just an opinion

1 Like