LLM在推理时,基于已经输出的句子,从词汇表中预测下一个词。预测时,为词汇表中的每个词分配一个概率,通过采样得到预测输出,如图1所示。例如,模型在输出"My name is"后,仅有0.62的概率输出自己的名字"Tang",即使我们在提示词中告诉了模型自己的名字是"Tang",模型也有0.38的概率输出别的名字。
It’s the season to treat yourself to tricks and treats, boos and booze, and plenty of spooky fun. From now until November 2, Jing An Shangri-La, Shanghai presents an autumn pop-up Enchanted Bar, ...
Now bigger, scarier, and more death-defying than ever before, Shanghai's world infamous, oft-replicated-never-duplicated original Halloween Tribute Show is happening on two nights, November 1 and ...