How to do gpt model fine tune well?

API
1

Hey Guys!

After fine-tuning with the following data, no matter what words I enter in the prompt,
only those words(food, fashion, travle…) will be answered. There are less than 300 datas!!

example.jsonl
{"role": "user", "lunch"}, {"role": "assistant", "content": "food"}
{"role": "user", "Cooking is so easy"}, {"role": "assistant", "content": "food"}
{"role": "user", "I'm hungry"}, {"role": "assistant", "content": "food"}
...
{"role": "user", "cool style"}, {"role": "assistant", "content": "fashion"}
{"role": "user", "It's so cold day"}, {"role": "assistant", "content": "fashion"}
{"role": "user", "Give me some pants?"}, {"role": "assistant", "content": "fashion"}
...
{"role": "user", "travle..."}, {"role": "assistant", "content": "travle"}
...

I didn’t want it to be fine-tuned like this.
I want…
For example
‘After you answer like that, please indicate the accuracy of your answer additionally.’
then the answer is ‘food.’ 85’
not just only food.

Can you give me some advice?

2

To achieve responses like food - 85%, follow these steps:

1 Structured Dataset: Organise your data with questions, answers, and associated accuracy percentages.

2 Fine-Tuning: Train your model using this structured dataset to generate responses that include accuracy indicators.

3 Response Formatting: In your code add accuracy percentages to the model’s answers based on the dataset.

Like this:

{"question": "What's your favorite food?", "answer": "food", "accuracy": 85}
{"question": "Tell me about fashion trends.", "answer": "fashion", "accuracy": 90}
{"question": "Best travel destinations?", "answer": "travel", "accuracy": 88}
3

thank you for your reply

If I ask a question that isn’t in the finetunning data, I want to be able to accurately indicate whether it’s about food or travel.

If the questions in the finetunning data are exactly the same, the accuracy will be 100%.