Hi,
I’m wondering why in the lab we do save and reload the weights of the tokenizer after training. Did PEFT change the weights of the tokenizer ? I thought it would only affect the queries/values part.
Hello @Mordokkai
Because Parameter-Efficient Fine-Tuning (PEFT) methods enable efficient adaptation of large pretrained models to various downstream applications by only fine-tuning a small number of (extra) model parameters instead of all the model’s parameters.
We reload the weights of saved model and freeze the part of the model we do not want to use, and further fine tuning only the parameters we want to get a better llm model. This significantly decreases the computational and storage costs.
Regards
DP
Hi @Deepti_Prasad ,
Thanks for your answer but my question was concerning the tokenizer part especially. In the lab we use Lora to finetune, but I thought Lora only concerned the attention part of the llm (queries, values), I didn’t know it would also concern the tokenizer
It is all related, what if the peft doesn’t want to use all the tokenizers and only selected tokenizer and fine tune the llm.