What is the right way to structure `input` and `label` while fine-tuning decoder only model

I am working on fine-tuning a Llama(decoder only) model on SQuAD Dataset, and wondering what is the right way

to structure inputs and labels.

Option1:

input: “SOS context:{context} question:{question} answer: pad …. pad”

label: “{answer} EOS pad pad … pad”

Option2:

input: “SOS context:{context} question:{question} answer: pad …. pad”

label: “pad … pad {answer} EOS pad … pad”

Option3:

input: “SOS context:{context} question:{question} answer: {answer} pad …. pad”

label: “context:{context} question:{question} answer:{answer} EOS pad …. pad”