#API Is Way Worst at Classification than Browser-Based ChatGPT (using text-davinci-003)

7 messages · Page 1 of 1 (latest)

twilit crater
#

Parameters:
'model' => 'text-davinci-003',
'prompt' => $Prompt,
"top_p" => 1.0,
'max_tokens' => 450,
'frequency_penalty' => 0,
'presence_penalty' => 0,

I'm playing around with using GPT to detect form spam. I tested my idea in the chatgpt browser version and it worked great, but I can't replicate the results via OpenAI API. For example some obvious spam pasted as a prompt into browser gives me:
"Spam

This form submission appears to be spam as it includes multiple suspicious links promoting pharmaceutical products. The links are unrelated to the form submission and are likely an attempt to advertise or sell products."

But the response to the identical prompt from API is stuff like "Not Spam.\nThe form fields appear to be genuine and relevant"

I have tried to vary the top_p and temperature parameters without much success (it changes the reason for it not being spam, but still claims it's not spam).

Is the model we get via API not the same as the browser version? Is there any way to set the parameters to match the browser?

dusk gull
#

naturally, chatGPT is newer, in fact it wasn't even released yet

twilit crater
#

so there's no way to access the chatgpt system through API?

crisp sorrel
leaden wadi
#

I just faced the same problem when trying to detect inappropriate words and I also found that the Moderations API doesn't perform really well when compared to chatGPT and Playgrounds. If I figure something out I'll share it here.

vestal timber
#

You could also play around with best of