GPT ignoring simple instruction

API
1

Hello,

I’m having an issue with prompting GPT, it ignores the simple instruction, at least to me, it seems like an easy instruction.

image
image1282×363 33.5 KB

As you can see from the screenshot, it keeps wrapping the sentence with double quotes against my instruction.

Model: gpt-4-0613
Temperature: 1
Max tokens: 4096
Top P: 1
Frequency penalty: 0
Presence penalty: 0

Any suggestion would be appreciated.
Thanks.

2

I’d drop a quick example in the system message…and remove “Sentence should not be wrapped with double quotes…”

Add this to your system prompt…

Example:

USER: Some sentence
ASSISTANT: Headline without quotes…

3

Thanks for the reply, but this did not help.

4

What’s your system prompt that you are using currently?

5

Hi @PaulBellow ,

I followed your instruction by giving it an example first and then asking to rewrite the next input:

$prompt = [
	[
		'role' => 'system',
		'content' => 'You are a content writer for my blog. Write the post title based on the input.',
	],
	[
		'role' => 'user',
		'content' => 'Linux Kernel 6.6 Officially Confirmed as an LTS',
	],
	[
		'role' => 'assistant',
		'content' => 'Confirmation of Linux Kernel 6.6 as an Official LTS Version',
	],
        [
		'role' => 'user',
		'content' => $newTitleToRewrite,
	],
];

Most of the time the new title is good, but there are still cases where whole sentence is wrapped in double quotes "

An alternative solution would be to just trim the quotes programmatically once I get the response from API.