Number of characters generated in zero-shot vs. few-shot (davinci-instruct vs davinci)

I’m doing some testing for copywriting with davinci-instruct-beta only with some prompt guidelines of what to do at the prompt. I narrow down the results to the number of tokens I want GPT-3 to generate (characters // 4) and it works pretty well on average in terms of the number of characters generated and the quality of the results.

The problem comes when I transfer this same case to the davinci engine with a fixed prompt and add a few previous examples (few-shot learning). I see that the characters of the results are much less than the ones I specify as max_tokens.

I understand that the examples added obviously affect the results, but do they also affect the size of the generated text? Can anyone think of a way to control that from the prompt to “guarantee” an approximate number of generated characters (as it happens with davinci-instruct)?

As an example of prompts would be something like the Dynamic Examples case of this web:


1 Like

If Davinci Instruct gives you exactly what you need, just curious, why do you need to use Davinci then?

The use of davinci is to be able to structure the data in the examples for few-shot learning.
If you look at the gptprompts example I posted, there are entries like this one:

      "company":"Jimbo's Backpacks",
      "audience":"School kids",
      "tone": "Urgent",
      "promo": "15% Off All Backpacks",
      "keywords": "Customer Service, Local, Backpacks",
      "background":"Jimbo's backpacks has sold backpacks for years. Our backpacks are locally sourced. With school approaching, everyone wants a bike. Great customer service.",
      "AdSense Headline": "Great backpacks for School | Coolest Backpacks in Town | While supplies last",
      "Final Description": "Hey kids, need the coolest backpack that fits you well? Our professional backpack pickers can help you prepare for the new school year. Backpacks are flying off the shelves, get yours before they're gone.",

All the examples I see of this type of training for these situations are with davinci engine and I have not seen others with davinci-instruct to see how to “tell the model” what to do in this way.
If anyone has an example of how the prompt for davinci-instruct could be structured, please let me know.