Finetuned and aligned models

[shanefeng123]

Hi,

I'm following your code to fine-tune and align the models. However, this is going to take a long time. I'm wondering if you can upload the models you used to HuggingFace so that I can directly use them to find the safety neurons, and skip all the training steps? Since you're using IA3 to fine-tune and align the models, these should be very small in sizes.

[chenjianhuii]

Since I have graduated from THUKEG, I cannot access the checkpoints now. Besides, you don't always need to go through the SFT stage if you have a model that is not safe enough, you can just perform DPO on it.