GPT-3 responding with blank text

API
1

I’m trying to set up the GPT-3 API on Bubble, but keep getting an empty text response in return to my calls.

I’ve added my key to the header and the content-type is set to application/json. I’m passing the following sample JSON (from the docs) via a POST to https://api.openai.com/v1/completions

{
  "model": "text-davinci-003",
  "prompt": "Say this is a test",
  "max_tokens": 7,
  "temperature": 0,
  "top_p": 1,
  "n": 1,
  "stream": false,
  "logprobs": null,
  "stop": "\n"
}

but when I initialize the call this is the response I get with a blank text response

Returned Values:

{
    "id": "cmpl-6......YXhL4y",
    "object": "text_completion",
    "created": 1673992522,
    "model": "text-davinci-003",
    "choices": [
        {
            "text": "",
            "index": 0,
            "logprobs": null,
            "finish_reason": "stop"
        }
    ],
    "usage": {
        "prompt_tokens": 5,
        "total_tokens": 5
    }
}

Does anyone have any suggestions? I’ve tried reaching out to OpenAI support but I guess they are super backed up.

1 Like
2

Just a guess but your tokens are set to 7, try going to 100. Also, if that doesn’t work, your stop is ‘\n’ so put a \n at the end of the prompt.

1 Like
3

You might consider starting with the defaults and see what happens, and then refine your API query from there:

{
  "model": "text-davinci-003",
  "prompt": "Say this is a test",
}

Or maybe, this:

{
  "model": "text-davinci-003",
  "prompt": "Say this is a test",
  "max_tokens": 1024,
}

Then, you can receive the response and refine your query from there.

HTH.

4

@adrianneestone did you end up figuring this out? I’m having the same issue, but I’ve already increased my max tokens to 2048.

5

It’s probably this…

You’re telling it to stop with a line break…

And when it starts the output, it starts with… a line-break…

Hope this helps…

6

I ended up getting it to work with "stop": ["\\n"]

I was getting super frustrated though because the code I was using came from the “Create Completion” dang example in the docs!

2 Likes
7

I was doing some more experimenting, and it turns out that slight edits to my prompt will result in one of two scenarios of incompletion. To give a little more context, I’m providing the same prompt multiple times, but subbing in the text that I want GPT3 to deal with. In some iterations of my prompt, I get the scenario that we’ve been talking about here, where all predictions are empty. However, for some iterations of my prompt, I get a few documents with predictions, and many others that are also null. Do folks have thoughts about why that may be? I haven’t changed the stop character yet.

Examples of my prompt for each case:

prompt = ('Extract the biological relationships from '
'the following text as (Subject, Predicate, Object) triples, and '
'include the index of the sentence from which the triple is extracted: '
f'{abstract_txt}.\nExample relationship triple:\n("Protein 1", '
'"regulates", "Protein 2", "Sentence 0")')

This prompt gets some responses.

This prompt, on the other hand, gets none:

prompt = ('Extract the biological relationships from '
'the following text as (Subject, Predicate, Object) triples, and '
'include the index of the sentence from which the triple is '
'extracted: '
f'{abstract_txt}.\nExample relationship triple:\n("Protein 1", '
'"regulates", "Protein 2"), "Sentence 0"')

In all cases, the {abstract_txt} variable is a list with one element, which is a string containing the text of a scientific abstract.

I’m wondering if there’s something else going on that just happens to correlate with me changing my prompts, or if it’s the prompts themselves? Why does such a small change result in no output, and why don’t I get output for all abstracts in either case?

EDIT: formatting

8

I had the same problem and your solution solved mine as well… Any idea why this seemingly random stop signal solved it?

9
1 Like
10

Thanks for your suggestion …It worked with “stop”: [“\n”]

1 Like
11

This is because you simply have removed the newline from your STOP because:

["\\n"]

Is not a valid newline char because when you escape the newline as you have done, there is no longer a newline stop.

All you have done @adrianneestone is, by accident, removed the newline stop and your stop is now the literal “backslash n” instead of a newline.

In other words:

This is a newline stop

["\n"]

This is NOT a newline stop (because you escaped the newline)

["\\n"]

HTH

:slight_smile:

Appendix

From the Ruby console you can see clearly that when you put a backslash in front of a newline, the newline is escaped and is no longer a newline:

irb(main):008:0> a = "hello world\\n"
=> "hello world\\n"
irb(main):009:0> puts a
hello world\n

irb(main):010:0> a = "hello world\n"
=> "hello world\n"
irb(main):011:0> puts a
hello world