I’m trying to use few-shot learning to try and get GPT to generate answer to a question in a specific manner.
However, the generation for a question completely unrelated to the sample questions is still being influenced by the sample answers they have.
Anyone who has encountered this and has a prompt tweak that might help sort this out
Cheers,
Ud
Are you talking about fine-tuning a particular base model @udm17 ?
Not fine tuning. Using samples along with a prompt to generate the answer for a question. Sadly can’t share more info as it is for a propriety software.
What settings are you using? temp, freq_penalty, etc
Temperature - 0
Freq Penalty - 0
Top P - 1
Presence Penalty - 0
In my experience, the lower you go with temperature, the more likely it is to overfit/repeat… If you can’t raise the temperature, try moving frequency_penalty up a bit… but slowly… 0.05 at a time maybe?
Good luck!
Cheers Paul.
This is something I have been tinkering with a bit, hopefully can find the sweet spot soon. I want the answer to be deterministic to a certain extant, so have been using low temperature and that definitely has led to overfitting.
Base model I’m using currently is davinci003 and has recently started using the gpt3.5 turbo. Am not too sure whether I should use a fine-tuned davinci would solve my problem.
Fine-tuning is the best OpenAi tool available for better model fitting.
HTH
