Questions on peft_lora_seq2seq example

[siddhsql]

hello - i have some questions re: this tutorial:

1. what is the purpose of this line:

labels[labels == tokenizer.pad_token_id] = -100

2. and how does the batch_size make a difference in practice? e.g., should we use as large a batch size as we can given memory limitations? what should be the optimal value? (trial and error is not an answer)

3. is max_length = 128 the maximum length of a given sentence?

4. also will the steps in this tutorial work for any seq2seq model?

thanks a lot.