How to reduce messy sentences?

Hi!

I’m getting paragraphs like this: Screenshot by Lightshot

Any chance to get rid of them?

Cheers

What’s your prompt and settings look like?

From what little info I have to go on, I would say it’s temperature or frequency_penalty set too high.

Is this a fine-tuned model or a standard model?

1 Like

It’s basically: create outline for X topic, after it does i instruct it to generate paragraph for each.

Temp is 0.5, but freqquency_ppenalty is at 2… but i find that higher they are, more detailed the output is :confused: it’s still good at 0.5 each.

What settings would you advise for temp, f_penalty and p_penalty, considering my prompt?

It’s fine-tuned. They are a few variables and instructions.

1 Like

I’d say your frequency_penalty is way too high. It’s a very sensitive setting. What I think happens is that you’re telling it under no circumstances repeat anything, so it’s getting “smart” and taking out spaces so the tokens don’t match up…

Try taking frequency_penalty down to 0.5 or maybe even lower. I tend to not go over 0.08 or 0.12 a lot of times. I’d not introduce p_penalty into the equation at all - ie don’t use it.

The problem is that with a low temp you tend to get a lot of repeats too. So, try to also take the temp above .7 maybe .73 or so to test. I’d actually start playing with the knobs and test, but I’m betting your super-high f_penalty is mucking it up.

Let us know if it helps. Thanks, and good luck!

ETA: Although frequency_penalty goes up to 2 on the playground, if you’re using the API, I think it expects a number between 0 and 1… Can anyone confirm? @staff

2 Likes