Guidance in fine tuning with Question/Answer

API
1

Hi,
I was exploring OpenAI and my goal was to train the model with some specific set of question answers. I prepared the dataset in the following manner

{"prompt": "What are the services XYZ provides?", "completion": "Hi, XYZ provides Web and mobile application development services, Outbound call center solutions, and IVR services."} ...

I trained the model, then used the model with prompt

openai api completions.create -m my-generated-model-xyz123 -p "What are the services XYZ provides?"

Unfortunately, I am not getting the desired result. I’m sure, I am doing something wrong. I tried ada and davinchi for testing but no luck. Need your guidance to find out my mistake and fix the same.

Thanks in advanced.

2

You should try using a Q:/A: format for your tuning examples, like below:

{“prompt”: “Q: What are the services XYZ provides?\nA:”, “completion”: " Hi, XYZ provides Web and mobile application development services, Outbound call center solutions, and IVR services."}

Then implement this same schema in your inference calls to GPT3

3

It improved the result a bit. However, I’m also trying some more approaches to get the best output.

Thanks for your suggestion.