Interesting
#1
by Trilogix1 - opened
According to my tests it keep the accuracy. can you do all qwens 3.5?
This method is successful, good job.
I actually have finetuned and improved on the performance and response structure. After a bit more rigorous testing, I realize the ablation strength of 0.35 corrupts the output more often than not. So I will be uploading another variant which is finetuned on uncensored text data w/ lora, and I hope you enjoy how much better it is now.
I will try to do other variants too :D, but I am resource-constrained :(
Well according to my tests (not bench) this one perform better preserving accuracy with a low refusal rate.
I tries the other ones but this is the best.
Would you mind sharing which method you used, which scripts and repo?