Fine tuned model cannot do basic max min

Prompting
1

I’ve fine tuned the ada model with 100,000 examples resembling:
{“prompt”:" price of Apple iPhone 11 Pro 256GB carrier Other condition GoodDevice date May 2021 \n\n###\n\n",“completion”:" Other 256GB GoodDevice 400\n"}
{“prompt”:" price of Apple iPhone 11 Pro 64GB carrier Other condition GoodDevice date May 2021 \n\n###\n\n",“completion”:" Other 64GB GoodDevice 340\n"}
{“prompt”:" price of Apple iPhone 11 Pro Max 256GB carrier Other condition GoodDevice date May 2021 \n\n###\n\n",“completion”:" Other 256GB GoodDevice 455\n"}
{“prompt”:" price of Apple iPhone 11 Pro Max 64GB carrier Verizon condition GoodDevice date May 2021 \n\n###\n\n",“completion”:" Verizon 64GB GoodDevice 385\n"}

I can get one of the fine tuned values when I prompt like:
price of Apple iPhone 11 Pro Max 64GB condition GoodDevice

However, I’m looking for the data I fine tuned with to be interpreted for more complex analysis. However, the model cannot even return minimum and maximum values with prompt:
minimum price of Apple iPhone 11 Pro Max 64GB condition GoodDevice
maximum price of Apple iPhone 11 Pro Max 64GB condition GoodDevice

it is as though the fine tuning turns off any understanding of the tokens in the fine tuned examples. I would expect that it would identify the tokens and form a response using the tokens independently. Do I have to actually ad a prompt and completion for analysis, such as min, max, diff, delta,…

2

Use Curie with about 1k-2k examples, not more than this. Try to prompt everything that impact on completions.

It is only a suggestion based on my experience, no guarantee that it will work as you expect.