How to fix my fine tuned models responses

API
1

Here are some responses from my model that I fine tuned.

what if I dive a red car and crash it when I’m drunk?-> The insurance will not cover any damage or theft if you are intoxicated or if you violate the terms of the rental agreement.

The above is a good response. The below is eh.

The airline cancelled my flight and I’m going to miss my connection in singapore. What should I do?->If your flight is cancelled and you’re going to miss a connection, consult the ‘Additional expenses’ section of your policy for information on what to do. It may include making a claim for additional accommodation and transport expenses or taking advantage of any alternative travel options.

Another good but bad response:

What if i damage the rental car?-> The Rental Car Insurance Excess should cover the expenses as long as the rental car gets damaged while in your control during your period of cover, and you are in compliance with all the conditions.

These examples have the correct information entirely, however it hasn’t put it in the right order per say, and has added some things I didn’t want.

My question is this, do I just need to run another epoch and clean the data, or is there something else I can do? Or would the System prompt fix this?

I’m working with exactly 2815 prompt and completion pairs for this. Trying to breakdown an insurance policy, as you can see.

2

Hi,

If you have a company corpus of data and you feed that into the model as fine tuning, it will learn the style of the interactions, it will not learn the interactions themselves. For example, if you were to show it thousands of books an author has written, it will learn “how” the author wrote, not what the author wrote.

With this in mind, possibly you are using the wrong tool for the task, the alternative is embeddings and in those you can feed it your company documentation in sections and then perform a semantic similarity search on the data to retrieve relevant sections which can then be included as context in a prompt containing the users query, see OpenAI Platform