Hi,
We fine-tuned a davinci model using articles on our website. But, when we try to have a conversation with the model, it keeps repeating information as answer to a questions. Like ”answer phrase 1, phrase 2, phrase 3, then it keeps repeating phrase 3, phrase 4, phrase 3 and so on”.
Did you encounter a behaviour like that? Is it because of the prompt? Because of the training data? Data is good. The answer is pretty ok, but the repetition is kills the effect.
Thank you.
What settings and model are you using?
We fine tuned a model using davinci.
As for the settings:
temperature=0.5,
max_tokens=300,
top_p=1,
frequency_penalty=0,
presence_penalty=0
Hi @gealpop
Based on your description of the issue, setting frequency_penalty to a suitable value could be a possible solution.
More info here.
This could also be related to how you set up your separator and stop when you trained your fine-tuned model.
Please post back a few lines of your JSONL file so we can advise further, @gealpop .
Also, please provide your fine-tuning params when you fine-tuned.
Thanks!
