Fine tune model with some incorrect training data in large data set

I created the fine tune model with the 1000 training data in this training data by mistake I added the 30 training data with incorrect guidance.

so this affects the out put of the created fine-tune mode.

if I created a new model then it would be costly to give a price for the whole fine-tuned training data model.

Is this any way to remove these 30 training data from the existing fine-tuned model?

Not currently, finetuning cannot be removed after it has been encoded, so you will have to create a new set with the incorrect elements removed.


If the 3% of your training data makes any difference at all, you can make a continuing fine tune model just on a small training file of identical input and corrected outputs as your training data, run that for twice the epochs of your original job (or more), and see if that fixes that domain of answering.

Could you please let me know how to Continue fine-tune?

The first time that you created a fine tune job and you specified the model was gpt-3.5-turbo?

The second time, specify the model is your fine-tune model name.

1 Like