Can only get a limited number of logprobs from completions?

[yangkevin2]

Hi,

I am specifying that I want the logprobs of the top 100 tokens but the API seems to give me at most 5:

(Pdb) response = openai.Completion.create(engine='ada', prompt='test', max_tokens=1, logprobs=100)
(Pdb) response
<OpenAIObject text_completion id=cmpl-3qtEQ4P8iNmTHqV8Eh4Wq59vzqnbm at 0x7fcf271ebb30> JSON: {
  "choices": [
    {
      "finish_reason": "length",
      "index": 0,
      "logprobs": {
        "text_offset": [
          4
        ],
        "token_logprobs": [
          -10.029023
        ],
        "tokens": [
          " goal"
        ],
        "top_logprobs": [
          {
            ",": -3.721809,
            "-": -2.6293674,
            ".": -2.6775126,
            "/": -2.527664,
            "_": -2.10408
          }
        ]
      },
      "text": " goal"
    }
  ],
  "created": 1633734378,
  "id": "cmpl-3qtEQ4P8iNmTHqV8Eh4Wq59vzqnbm",
  "model": "ada:2020-05-03",
  "object": "text_completion"
}

Is there any way I can get around this?

[BigSalmon2]

openai.api_key = "[X]"
response = openai.Completion.create(engine="davinci-codex",
                                    prompt="Hello",
                                    logprobs=100,
                                    max_tokens=1)
print(response['choices'][0]['logprobs']['top_logprobs'][0])```

[yangkevin2]

Er, was this supposed to be a fix? The response still only has 5 logprobs when I run that command.

[kimyoonduk]

Having the same issue with logprobs being hardcoded to 5, even as I send POST requests to the API endpoint via curl.
Perhaps this is a serverside problem? Definitely would like to hear updates on this!

[hallacy]

I'm a bit late here. There default was recently changed to 5 as mentioned here: https://beta.openai.com/docs/api-reference/completions/create#completions/create-logprobs. If you'd like more, please reach out to support@openai.com