Do I have to use davinci to get meaningful results?

Prompt Assistance
1

Hi all,

I’m still fumbling around with my fintech virtual assistant! In part of the tool, the user is asked if they want to work through additional methods of saving money in retirement or adding to their pension pot. I have successfully used the following prompt with davinci:

A user is asked how he wants to spend less money in retirement, or add additional funds to his retirement pot. From a choice of 5 options which are:

  1. Reduce essential expenses,
  2. Reduce goals,
  3. Generate additional income,
  4. Use property,
  5. Include an expected windfall,
    the user says
    “Free up some equity”.
    Which option did he mean?

To which I get the response

“4. Use property”, which is correct.

“work part-time” gives me “3. Generate additional income”
“Include my inheritance” gives me “5. Include an expected windfall”
“Spend less money” gives me “The user meant to say Reduce essential expenses.”

So davinci works beautifully and gives me a response I can work with very easily without asking the user to utter a specific phrase. It’s perfect.

Until I use a model other than davinci. Then I get completely different (and incorrect) results.

I don’t mind using davinci except that it is so slow, the user starts repeating their answers because they think they haven’t been heard.

How can I either speed up davinci’s turnaround or else phrase the prompt so that I get more meaningful responses from one of the faster models? If that is even possible, of course!

Thanks so much,

2

I believe that a fine tuned Davinci model is approx 300% faster than using raw Davinci 2.
And a fine tuned Curie model is significantly faster than a fine tuned Davinci model.

I would approach this by fine tuning Curie on a large dataset (at least 1000 examples) and evaluating the quality of responses (it should be more than capable if trained on a high quality dataset).

If the generated output from a fine tuned Curie does not reach your expectations, I would proceed by training a Davinci model on the same dataset.

3

Thanks @jeffinbournemouth, I guess I need to do some research into fine tuning then!

By the way, you’re just up the road from me! I’m in Christchurch :upside_down_face:

4

Hi George, good to meet you!

I live in Muscliff, near to the nature reserve.

Yes, I think fine tuning Curie may offer the best combination of speed & quality, and affordability.
The more examples you include in the dataset, the better quality you will achieve.

There are several sources of info on openai.com for training a custom model.

Good luck!