How to reduce the cost of model training with davinci?

API
1

I’m using davinci to fine-tune model for my project, somehow i have charged $6.37 for 2,000 lines of data model training only.

I have trained data only the testing purposes, in terms of the real data, which ca be reached more than 100K, then it will cost me enough amount to stop the project.

I just want a brief procedure to have best practices to reduce the cost of training data, or write smart prompt and completion as well.

Thanks!

2

Can you share some examples of your training data and what you are trying to achieve?

3

Please find below examples of my prompt and completion.

{“prompt”:“Which countries are supported by the Awin network?\n\n###\n\n”,“completion”:" The Awin network supports the following countries: UK, Benelux, Brazil, DACH, France, Italy, Nordics, Spain, Eastern Europe, US & Canada.\n"}

{“prompt”:“What is the network strength of Awin?\n\n###\n\n”,“completion”:" Awin has a network strength of 270,000 Publishers and 25,000 Advertisers.\n"}

4

Thanks. For this kind of fact-based FAQ bot, it is probably worth trying to use semantic search instead of fine-tuning. Here is a tutorial: OpenAI API

5

Thanks for the suggestion, but I don’t have only FAQs, I have performance-based analysis, multiple data sources, including website page structure checks, Content SOP checks, and many more, and combined them in a single place. plus I have Roman language as well.

For that reason, I’ve chosen to fine-tune, please suggest something regarding this scenario ?

6

Sure. It is hard to determine without more details. Can you give some examples that are not FAQs?

7

Actually, I’ve just started my development, and I’ve faced a cost higher than expected, to reduce it, do you have any suggestions?

8

Without more examples, it’s still not clear why you need to do fine-tuning, or if fine-tuning would work. For example, what is “performance-based analysis”? Why can’t you use chatgpt 3.5 to do it? Why can’t you fine-tune Curie to do it?

9

Generally, if you want to give the model new or specific knowledge you need to use embeddings.

Generally, if you want to teach the model how to answer a question (format, structure, etc) you would use fine tuning.

10

it sounds like more of an embedding’s solution than fine tuning. Is the data you are giving it static or might it ever change?

12

Sorry I misread. Not sure about your use case, but fine tuning should not be necessary unless, you invent something entirely of your own which GPT is not aware of.