一些细节的关于paper和代码的问题
LeeSureman opened this issue · comments
1.看paper你们似乎是先找的label mapping再找的template，但是在找label mapping的时候也需要template，请问这个时候的template是手动定的么？
2.请问你们的代码应该用多少卡跑？好像现在的代码对应的总batch会随着卡数的增加而增加，从而导致训练过程的不同
3.你们的readme里提到了跑sst-2的命令，在“Our final results (LM-BFF) take prompt-based fine-tuning with demonstrations, filtering and automatic template, for example:”后面，但是后面给的这个命令似乎没有用到你们搜到的label mapping，只用了template，而mapping就是manual的，请问就是这样的嘛
Hi, thanks for your interest in our work!
About automatic search, template search starts from manual label words, and label word search starts from manual templates. In either method, we need to start from something manual.
About GPU cards, we always use one card. If you are running it on a multi-GPU machine and only want to use one card, you should set CUDA_VISIBLE_DEVICES to the desired card ID.
We have mentioned in our paper that we found automatic template (+ manual label words) works better than automatic label words, so we follow this method in our final LM-BFF. So yes to your last question.
哇，感谢你的及时回复！太棒了