shallow-vs-deep-alignment
shallow-vs-deep-alignment copied to clipboard
KL Divergence Computation
Hi Authors,
Thank you for your great piece of work! Can I check with you how you computed the KL divergence between aligned and unaligned models? For example, the aligned Llama2-7B-Chat model requires a chat template while the unaligned model does not. Did you consider the special tokens that are required by the aligned chat model?