When I switch my organization, I find that the same model(e.g. text-davinci-003) responses different text. Not just a little different, but very odd.
My prompt is as follow:
Human: Hello!
AI: Hi.
Human: May I be your best friend, please?
AI:
other params:
model='text-davinci-003', stop=['AI:', 'Human:'], temperature=0.9, max_tokens=150, top_p=1, frequency_penalty=0.0, presence_penalty=0.6.
In the 1st organization, the text of response:
Sure! I’m always happy to have a new friend.
In the 2nd organization, the text:
\n\nMaster: My boy Makumo, your Time Machine is just going to end in your death.\n\nMakumo's Reaction: Please stop being careless, please\uff01\n\n
It takes up all the tokens, but doesn’t have anything to do with the prompt. Does anyone know why?
tianlujunr:
Hi @tianlujunr
Not sure what it up on your end and in your code, but on my end, it word fine, FWIW:
Setup
Completion
HTH
udm17
March 9, 2023, 5:32am
3
It is highly unlikely that the organisation has anything to do with the output of the completions. At 0.9 temperature, it might simply be the randomness and creativity coming into play.
I tried again with temperature=0.0. It’s still weird.
1st response(abnormal, the text is stuffed with spaces, notice the finish_reason):
{
"choices": [
{
"finish_reason": "length",
"index": 0,
"logprobs": null,
"text": " "
}
],
"created": ...,
"id": "...",
"model": "text-davinci-003",
"object": "text_completion",
"usage": {
"completion_tokens": 150,
"prompt_tokens": 38,
"total_tokens": 188
}
}
2nd response:
{
"choices": [
{
"finish_reason": "stop",
"index": 0,
"logprobs": null,
"text": "\nYes, of course you can be my best friend."
}
],
"created": ...,
"id": "...",
"model": "text-davinci-003",
"object": "text_completion",
"usage": {
"completion_tokens": 12,
"prompt_tokens": 25,
"total_tokens": 37
}
}
So do you have interest to be a Member of my abnormal organiztion and have a try?
udm17
March 9, 2023, 9:38am
6
Was the return of the first prompt just spaces just spaces at temperature 0 ?.
Yep… temperature=0, max_tokens=150… all the params are the same, except the openai.organization = “org-Lvg…”. That’s why I am confused.
tianlujunr:
hat’s why I am confused.
Just tried it again for you @tianlujunr .
Still works fine from here:
It’s not about the “Prompt” nor “Params”, but the “Organization”. When calling the api, you need an api_key, right? If you want to use the quota of your organization or company, you would have to join an organization. Look at the picture below:
I am a member of both of these two org. In the 1st one, everything goes well. In the 2nd one, the text of response looks wierd.
@ruby_coder Any idea?If interested, you could have a try. I’m glad to pick you up into my organization.