Fine-tuning theory
Hi, I am new to NLP. I would like to know more about theory aspect of your work. When you fine-tune the model, how to identify which layers are trained and which layers of LLM are frozen? Thank you so much in advance.
Hi, to be frank, I didn't look into each layer much, but from my understanding, the pretrained model already have some knowledge, so I fine-tune all the layers with the input format that I want with small learning rate.
However, as you pointed out, we might achieve better results if we freeze some layers of the encoder model. I think it is depends on experiment and select the best way to tune a model.
Hi, sorry for disturbing you again. I replicate your code in github to train my own model but I have a question about evaluating model cause I don't see you use any metrics to evaluate the efficiency of your model. Which metric should I use if I want to assess my model? Thank you so much in advance.
since this is a text generative task, I think you can use bleu score to evaluate model
you can check this for more information
https://huggingface.co/spaces/evaluate-metric/bleu