do full r1
that would be chad
Bro, that would be crazy... I have no idea how to get sufficient hardware🥲
what happens if you set scale factor to something ridiculous like 10 on your github repo to make these?
Well, the model will be fucked up, just infinitely generating rubbish. Scale factor >= 1.5 is the dangerous zone, and >= 2 will damage the model at serious risk.
BTW, I tried abliterating deepseek-r1-distill-qwen-7b, but my code didn't work. Refusal_dir was an array of nan. I have no idea what was wrong with my code 😶
One more thing, I am now on vacation, so my server in my office was shutdown. Considering my modest deposit, I can't rent a cloud GPU to figure out what exactly happened in abliteration process.
Well, the model will be fucked up, just infinitely generating rubbish. Scale factor >= 1.5 is the dangerous zone, and >= 2 will damage the model at serious risk.
BTW, I tried abliterating deepseek-r1-distill-qwen-7b, but my code didn't work. Refusal_dir was an array of
nan. I have no idea what was wrong with my code 😶
Tensor nan was fixed here