Finetuning LLM such as LLaMA2 to single task rather than multi task

105 views Asked by At

I am trying to fine tune vicuna 7b with a single task (training dataset size = 28k). When doing multi-task tuning, i only trained the model for 1 to 3 epochs and did not check the validation loss. However, since I want my model to be tuned for a single task, I would like to train it for more than 3 epochs and check the validation loss for early stopping.

When i tried fine tune the model using the same hyperparameter as for alpaca training (learning rate 2e-5), i found that evaluation lose increased after first epoch. I thought i should have trained my model more than 3 epochs for getting great performance for single task. However, it seems that my model is overfitting after first epoch.

Should I change the learning rate? Or what should I do?

0

There are 0 answers