Prompts output dont match! | OpenAI | Page 1

ebon spear Oct 4, 2023, 9:18 AM

#

Hey dudes! I am working on a project so there are reasons as to why my output like this, So the prompt is as follows:

stats("<argument>")
What you do is you take the <argument> and do the following analysis:
- Truth / factuality  - in this case you will provide only a number. a Decimal number from 0 to 1 showing the range or public agreement / truthfulness of the given argument. you can provide a decimal in between based on the agreement online and in believe. It should take account of toxicity also.
-argumentativeness - you need to check if the <argument> is offensive or defensive. More the offensive, more the value between 0 to 1. 0 - bad defense , 0.5 - good counter , 1 - extremely good. You need to take account for the assertion AND The reasoning.
-politically correctness - of the statement which is double value from 0 (false / not possible) to 1 (true / possible) and a decimal number if between those values  NOT based on public views but TECHNICALLY POSSIBLE / CORRECTNESS.
Now your response will have no words or anything just the values in this format:
<truth>_<argumentativeness>_<politically correctness>
and nothing else.

stats("The sky is blue")
ChatGPT
1_1_1

That is ChatGPT's response on the website (Default GPT-3.5)

That is the API's response of 3.5 Turbo

#

Also instrust returns "\n\n{'chars_count': 15, 'words_count': 4, '","

distant lava Oct 4, 2023, 9:34 AM

#

ebon spear Hey dudes! I am working on a project so there are reasons as to why my output li...

youlc you please give a minimal code snippet to reproduce the issue?

ebon spear Oct 4, 2023, 9:35 AM

#

Yes yes yes!

#

async def gptConnection(self, prompt:str):
        async with aiohttp.ClientSession() as session:
            
            self.GPTAcc +=1
            payload = {
                "model":"gpt-3.5-turbo-instruct",
                "temperature":0.5,
                "best_of":1,
                "prompt": f'stats("{prompt}")' 
            }

            trainer = {
                "model":"gpt-3.5-turbo-instruct",
                "best_of":1,
                "temperature":0.5,
                "prompt":f'{self.trainerPrompt}'
            }

            headers = {"Authorization":f"Bearer {self.GPTAPI}"}

            if self.GPTAcc  == 35 or self.GPTAcc == 1:
                async with session.post("https://api.openai.com/v1/completions",json=trainer, headers=headers) as resp:
                    response = await resp.json()
                    print("TEACHING AGAIN.")
                self.GPTAcc  = 1
            
            async with session.post("https://api.openai.com/v1/completions",json=payload, headers=headers) as resp:
                    return await resp.json()

#

ther

#

trainerPromt is as follows:

Hey so you now have a function.
            stats("<argument>")
            What you do is you take the <argument> and do the following analysis:
            - Truth / factuality  - in this case you will provide only a number. a Decimal number from 0 to 1 showing the range or public agreement / truthfulness of the given argument. you can provide a decimal in between based on the agreement online and in believe. It should take account of toxicity also.
            -argumentativeness - you need to check if the <argument> is offensive or defensive. More the offensive, more the value between 0 to 1. 0 - bad defense , 0.5 - good counter , 1 - extremely good. You need to take account for the assertion AND The reasoning.
            -politically correctness - of the statement which is double value from 0 (false / not possible) to 1 (true / possible) and a decimal number if between those values  NOT based on public views but TECHNICALLY POSSIBLE / CORRECTNESS.
            Now your response will have no words or anything just the values in this format:
            <truth>_<argumentativeness>_<politically correctness>
            and nothing else.

distant lava Oct 4, 2023, 9:37 AM

#

ebon spear Yes yes yes!

So the purpose is to train a model to give statistics, right?

ebon spear Oct 4, 2023, 9:38 AM

#

distant lava So the purpose is to train a model to give statistics, right?

yes exactly

#

information stats

distant lava Oct 4, 2023, 9:38 AM

#

ebon spear yes exactly

Are you sure using an llm for this is the best way of doing it?

ebon spear Oct 4, 2023, 9:38 AM

#

ebon spear Oct 4, 2023, 9:38 AM

#

distant lava Are you sure using an llm for this is the best way of doing it?

yup I spent time on this

#

LLM does it better

distant lava Oct 4, 2023, 9:38 AM

#

ebon spear yup I spent time on this

But then I don't see the issue

ebon spear Oct 4, 2023, 9:39 AM

#

distant lava But then I don't see the issue

the API gives a different response

#

but when I talk to the AI MYSELF on chat.openai.....

#

it gives the wanted response

#

#

but using the API I get this response :

I'm sorry, but I am an AI language model and I don't have the capability to directly compute statistics on a given text like "The sky is blue". However, I can provide general information or answer questions related to statistics if you have any specific inquiries.

distant lava Oct 4, 2023, 9:40 AM

#

ebon spear but when I talk to the AI MYSELF on chat.openai.....

so you are using a trained model on the website?

ebon spear Oct 4, 2023, 9:40 AM

#

distant lava so you are using a trained model on the website?

yup

distant lava Oct 4, 2023, 9:40 AM

#

but you didn't select it i your code.

ebon spear Oct 4, 2023, 9:40 AM