Few shot learning with gpt-4 - is it needed and what is best practice?

API
,
1

I’m creating a tutorial on the OpenAI API/gpt-4. Back when the model of choice for text creation was text-davinci-003, it was common to use a “few shot” approach to improve results. This would involve adding one or more examples to the prompt.

Is “few shot” still considered good practice with gpt-4? And if so, should the examples go in the system object of messages array:

const messages = [{
role: 'system',
content: 'This is my instruction, and here are some examples of what I want.
###
Example 1, 
###
example 2'
}]

or should they be presented as a conversation:

const messages = [{
role: 'system',
content: 'My instruction.
},
{
role: 'user',
content: 'example question'
},
{
role: 'assistant',
content: 'example desired response'
}]

There doesn’t seem to be anything in the docs about this, and maybe few shot is considered unnecessary with gpt-4.

2

Personally I keep the system prompt for persona definitions and global preferences, use Celsius, you are a helpful assistant, etc. I put my shots in the user prompt, you could try experimenting with the system prompt and see how that works for you though.

3

Thanks Foxabilo!
Just to clarify, when you say you put your shots in the user prompt do you mean you hard-code the start of the conversation and in the system prompt instruct the model to continue the conversation in the same style?

Or is your user prompt something like

{ role: 'user',
content: 'Write me a covering letter for a job with these criteria ${job_criteria} and make the letter similar in style to these examples: 
###
example 1: ${example1}
###
example 2: ${example2}'
}

Many thanks!

4

I will define a section at the start of my user prompt “###{shot_examples}### Using the examples in ### markers, please {Task_to_be_performed}”

1 Like