GPT-3.5-turbo not skiping a line when requested


I’m using GPT-3.5-turbo-1106,

But despite instructions to skip lines to create organized and easy to read paragraphs,

The model produces blocks like this:

Temperature is set in a way that the model would follow my instructions, and I’ve paid attention to limiting the tokens sent so that enough remains for completion.

Do you have any guess for why does it still look like this?


PS don’t worry the data is not confidential, it’s mock data.

How about just using css to format the output?

.display_text {
line-height: 200%;
1 Like

These paragraphs look very readable to me. What, exactly, do you mean by “skip lines?”

In general, my suggestion would be to not get into a fight with the model about something like this. Because if “skip lines” means what I think you mean in this context—at the end of each line of text put two newline characters so the output is double-spaced—that is something no model which exists today could possibly do.

The models speak in tokens not letters, words, sentences, paragraphs, or pages, and tokens can be of any imaginable number of different display-widths even if you’re using a monospaced font, if you’re not then there is zero possibility the model will ever know where one line of text ends and the next begins, nor should it or should it care—that’s the job of the display of the text generated. The model’s job should end once that is done.

If you really wanted to, you could send the text back through the model with another job to format the display of the text, but asking it to do it in a single go and on-the-fly is asking way too much.

Once you have the text, you can figure out how to display it to your liking.

1 Like

Thanks a lot. I’ll consider it, even if adding API calls in my workflow is to be considered thoughtfully. I appreciate your overall feedback, and it’s true that the level of readability is already correct.