Davinci Re-Fine-Tuned model predicts incorrect labels

Hello everyone,

This issue has to do with fine tune models for a text classification exercise. I read the documentation and followed it as it suggested. I was trying to refine tune an already existing model; However, after a couple of refinements, the model started predicting labels which I have not included in the training phase. I encountered this problem twice, so I will give the context in both experiments.

I also need to mention that the retraining dataset was much smaller than the original. I read that this is possible and the openai engineers suggest a reduction at the learning rate variable by a factor of 2 to 4.

1st case

I retrained a model using the labels " Ir" and " Re" (suffix separator: ā€™ ->') and it started predicting labels relevant to the context This one is not a major issue since one of the two is present.

2nd case

I retrained a model using the labels " false" and " true" (suffix separator: ā€˜\n\n###\n\nā€™) and it started predicting weird tokens (like \n). This is a major issue since none of the relevant labels is present.

You can see the first case on the left and the second on the right (sorry for using a single screenshot, I cannot upload two).

I need to underline that in the second case, a substantial amount of such incidents occurred. Also, in the second case, it was after the 2nd retrain while at the 1st case, it was after the 5th retrain. Furthermore, I used logprobs = 2 meaning that it is a binary classification and I need the log probabilities for 2 labels.

Any thoughts on this?

Thank you in advance!

Having exactly the same problem. The model is fine tuned according to the examples in documentation.

the predicted labels bear no connection to the labels used in the training data. It appears that the model uses its own predicted labels. Thoughts?