RRHF icon indicating copy to clipboard operation
RRHF copied to clipboard

bug 计算sft损失的时候

Open shyoulala opened this issue 7 months ago • 2 comments

计算sft损失的时候label和logits貌似没有shift,是我理解有问题吗? 应该是new_logits = logits[:,:-1,:]

image

shyoulala avatar Nov 24 '23 02:11 shyoulala