How to avoid answers like 'yes...' or 'no...' and force to expatiate with more related info

API
1

Ok, I’m trying to make a technical support Chat for our customers and have been training ChatGPT with fine-tuning.

I created a JSONL file with more than 10.000 lines using data from our Knowledge Base (from Wordpress).

I generated the json with a prompt like this:

Generate 30 questions and answers from this text. Use a neutral tone for both questions and answers. If necessary, you can end with “Get more info here {URL}”

The problem is that, for a question like this:

Can I export my data to CSV?

The answer is correct, but… extremely basic:

Yes, you can export your data to CSV

But I want to add HOW to do the export to the answer if the model has that information (and yes, it does know).

If I ask “How can I export my data?” the model will answer with the detailed instructions, but I want to avoid asking twice.

I would try to generate the fine-tuning with a more detailed prompt, trying to force the IA to add more info. Something like this:

Generate 30 questions and answers from this text. Use a neutral tone for both questions and answers. If necessary, you can end with “Get more info here {URL}”. Don’t limit to just give a short answer. If you know how to do what customer is asking for, include the detailed instructions in your response. If you don’t, include a link to XXXX

I need to know if this is the way or if I have to change some settings to give more detailed/related information to the questions.

2

Is not clear how you want to achieve your goal.

ChatGPT uses the turbo model, which is not possible to train.

1 Like
3

:question:

ChatGPI is not fine-tuneable as @georgei replied.

What are you talking about @santycg :question:

:slight_smile:

4

Sorry, I meant I was training the davinci model. I misunderstood the concepts.

I created some fine-tunes and I’m using the Completion mode with my custom model.

5

No worries. In software engineering, details are very important.

As for me, I do not expect a lanaguage model to reliable format data.

So, when I want data formatted from the model, I parse the output and then covert the response to CSV (outside the API call).

:slight_smile:

6

Ok, after some research, I think that fine-tuning is not the best method to train the IA for my purposes.

I habe tried a lot of variations, but the answers usually include not reliable information at all. Even if I ask for the different versions of our software, it answers with totally invented names :sweat_smile:

I will try with embeddings and see the difference.

2 Likes