Give me suggestions to create prompt for gpt-3.5-turbo-instruct

Hey, I am new at GPT Instruct model. I don’t have any knowledge on this model’s prompt. Can you guys help me to learn?

Let’s try to create a simple prompt. I am giving you 3.5-turbo prompt, Now you guys tell me how can I create the same in gpt-3.5-turbo-instruct.

The gpt-3.5-turbo prompt is :

{ "model": "gpt-3.5-turbo", "messages": [ { "role": "system", "content": "Act like a professional hockey player. You will be asked questions, Solve all the questions and response in JSON format with object result tag" }, { "role": "user", "content": "What is Hockey?" } ] }

It’s a simple prompt, My actual prompt is 4000 characters long with lots of complex task.

(preface: there is no meaning behind “object result tag”)

The first thing: The “instruct” AI has not been tuned to chat with you. The AI model does not simulate or recognize that it is a “you”. So you don’t talk to it directly.

Secondly, it still has strong completion behaviors of underlying base models. If you don’t put in strong delineation that part of the context is input, and part of the context is a place to put output, it will continue writing the input. This actually can work in your favor when you understand completion-style prompting.

Then there is the ability for multishot.

To answer your old prompt with refined quality, a playground preset that shows multiple techniques (also with some playground-specific tweaks)

https://platform.openai.com/playground/p/Z6UTuRUPIxtduD1tZ3dqONhk?model=gpt-3.5-turbo-instruct&mode=complete

(edit: link fixed to go directly to “complete” mode in the API Playground.)

With full access to the context, your “prompt” can put the AI right in the middle of the json format you want it to produce, where it must complete the json with additional information.

2 Likes

Okay so can you provide me a simple demo prompt or some resources for “instruct” model? Cuz I haven’t found any resources online to learn.

1 Like

You have an API account? Click the link above. See the text of prompting - all the text you see is “prompt”.

Then press “submit” to get your “what is hockey” answer.

Yup, I have an API account. And i already tried 30-40 prompt today. But the problem is “instruct” model misunderstanding my prompt and giving me wrong result. That’s why I thought maybe “instruct” model have some other mechanism and that I don’t know.

Here is a prompt that I tried.

Act like a professional reviews analysis, I will provide you review data in json format.

Here is data breakdown :

  1. “review” : This field is for user’s actual review. Contains feedback or suggestions or emotions or anything.
  2. “name” : Name of the user.

Predict the user’s mode by analysing “review” data. And give me the output in JSON format.

Use :

  1. “mode” - For user’s mode.
  2. “accuracy” - Accuracy of your prediction.

Here is the review:

{
“review” : “I am not happy with your service”,
“name” : “Avijit”
}

Now the problem is - “instruct” model automatically creating 2-3 more reviews and then giving me the output

You don’t have to say “act like” because you’re not trying to fool it into not being chatgpt.

Because the output isn’t containtered, you’ll need a stop token. Otherwise it will continue to complete the text that came before. You can see that my example created a container - and I have a stop token. etc

1 Like

Ohhhhhhhhh, Thank you for the information. :white_check_mark: :white_check_mark: :white_check_mark: :white_check_mark: :white_check_mark: