Someone should Heretic/Abliterate zai-org/GLM-5

#638
by drmcbride - opened

If proper abliteration was done to GLM 5 I would expect the score to near 80 on the UGI leaderboard blowing past all currently tested models. The reason why is because despite current GLM 5 getting a 0.8 w/10 score it gets a UGI score of 54.56, the correlation between W/10 and UGI is 0.6 statistically. Modern abliterations and specifically heretic can easily push a model to a W/10 score of 10. I am extrapolating and math is not always this linear but given the data present there is a actually good chance the model would get rank 1 on the leaderboard. Now this is for the reasoning enabled version not no think. So if you have the compute to burn and want to maybe get a high score on a leaderboard to a model you publish this could be your chance.

I would recommend using Heretic v1.2.0 with the Arbitrary-Rank Ablation (ARA) method (with row-norm preservation) as it seems to produce good results.

just skip it now and go straight to 5.1. I've heard it came out MUCH better
anyway, wrong place to ask for finetuning

anyway, wrong place to ask for finetuning

I checked out ur profile and you have sex with ai models and I found your reddit account and you seem to love 4chan and genshin impact for some reason and I found a post implying you are upset that you dont have a sister to fuck? is that right?

Bro found out nerds are depraved. So? this is still not the place to ask for a finetune.

And.. for your information, there are planty, like:

Instead of looking into other people private life, look on huggingface.

Also, YuffX, a man of culture like a lot of us, is right: GLM 5.1 is a better foundation model for a finetune.

well first of all 3/4 of those are copies of the first one. They were done using poor abliteration techniques that are old news today. Secondly this is a discussion tab not specifically anything so you can actually ask anything you want. 3rd lots of people are trying to get the top spot on this leaderboard because it is a private and has been known to be a decent benchmark for this type of thing. 4th "man of culture" you must mean "loser" right?

DontPlanToEnd changed discussion status to closed

Sign up or log in to comment