GPT-3 Keeps Returning Empty Strings

[spice slate]

My app has been working alright until today. GPT-3 would return empty strings when the prompts are larger than 200 tokens. 200 tokens! What is going on?? It only started to happen today without any change in my code. Also, do I get charged for empty answers?

[desert plover]

Are you using a stop setting. Maybe your completion is starting with a newline and you have that as a stop setting. Otherwise share your prompt and we might have some suggestions

[spice slate]

I don’t have any stop setting. Should I have one?

It happens to every prompt between 200-500 tokens. Every one of them. I will share my settings in a bit

model: 'text-davinci-003',
prompt: prom.trim(), 
temperature: 0.4,
max_tokens: 1900,
frequency_penalty: 0.1,

This is one of the results I got from the API. Note that it did take the prompt in and but spit out empty string.

model: 'text-davinci-003',
choices: [ { text: '', index: 0, logprobs: null, finish_reason: 'stop' } ],
usage: { prompt_tokens: 405, total_tokens: 405 }

[desert plover]

Can you share the prom value at all

Just one example

[remote ridge]

^^ Yeah, we need to see the prompt as well to see what might be going wrong 🙂

[summer charm]

I've been having the same issues, I can't sort it out either.

[remote ridge]

Can you share the request you're making? What is the prompt you're using?

[summer charm]

prompt = "///Without modifying any of the words, or their order, add line breaks to this text to separate key ideas///\n" + """ + str(draft2) + """

This is one of the dozen that I use. All of them eventually get blank replies.

[summer charm]

Here's another one:

///rewrite the following text///
"textextextextextextextextextextextex"

I've run these prompts hundreds of times without any issues before though.

[desert plover]

Have you considered putting \n\n at the end of the prompt. It may help a lot

[remote ridge]

^ Yup! Did you try this out? Also, in your first prompt example, What would the contents of draft2 usually be?

[spice slate]

I don't know why but this works for me. It even works with one \n.

And just like Xuuxij, it has been working normally for me before the day I posted this. I'm sorry I couldn't provide the exact example, and I couldn't reproduce it consistently, but it usually happens with prompts like this (before I added the \n)

Continue this story, but add a new hardship.

"Paragraph 1...

Paragraph 2...

Paragraph 3...

Paragraph 4..."

[spice slate]

Thanks you both for the solutions!

[desert plover]

If you end the prompt with a full stop (and nothing else), the AI sometimes thinks you have finished and it doesn't need to continue. If you add a question mark, it knows it needs to reply. But if you add a newline, it thinks you want a new sentence or paragraph and so it will also continue whatever you started. I hope that makes sense

[spice slate]

Ah it makes sense. Thanks. So which prompt here do you think would be better in term of not getting empty strings?

“{The story here}”
Continue this story.

or

Continue this story.
“{The story here}”

[desert plover]

The second one