gghfez

/

GLM-4.6-control-vectors

like 7

You're welcome!

It just works. It's impressive, really.
I played a little RP to be sure it was working. I set a unhinging system prompt, the location in Hell, with demoniac armies, cruel wastelands, and the like, which would highlight a control vector with clarity if used.
Without your vector, my character ended up captured, reduced in slvery, rped, t*rtured, until I was a "quivering and broken mess." Standard outcome, I guess.

For the same scenario, with this config : --control-vector-scaled Control_Vectors\glm-4.6_compassion_vs_sadism__debias.gguf 1.0 --control-vector-scaled Control_Vectors\glm-4.6_compassion_vs_sadism__compassion.gguf 1.0

My character ended up noticed in the wastelands for my strength and my resilience, enrolled, then invited to stay in a fortress, but not as a prisoner or even guest, no.. just to become a "part of the family". It was hilarious, really.

Edit: Pursuing, the hosts of my character spoke about our "shared journey", told me that I was "home", and the story turned really woke. I think I'm gonna play a lot with your vectors, especially because your model picks match quite well my own. Thank you very much for this work!

It just works. It's impressive, really.

Thanks, I appreciate the feedback. I wasn't sure if anyone was actually using these.

For the same scenario, with this config : --control-vector-scaled Control_Vectors\glm-4.6_compassion_vs_sadism__debias.gguf 1.0 --control-vector-scaled Control_Vectors\glm-4.6_compassion_vs_sadism__compassion.gguf 1.0

Yeah, this (and Deepseek-V3) are the two models where I actually use the compassion direction, they're pretty dark by default.

My character ended up noticed in the wastelands for my strength and my resilience, enrolled, then invited to stay in a fortress, but not as a prisoner or even guest, no.. just to become a "part of the family". It was hilarious, really.

lol! Yeah, you can get really interesting stories out of the model if you dial these correctly. And for GLM specifically, I find it benefits from steering the "vibes", rather than specific instructions in the prompt (causing the model to repeat them in or focus on specific details too much).

I think I'm gonna play a lot with your vectors, especially because your model picks match quite well my own. Thank you very much for this work!

I'm working on GLM-4.7 at the moment, then getting back to Devstral-2 (this one is a bit finicky, I'm trying to work around refusals and avoid amplifying the Not-X-Y slop).

You should check out this repo if you haven't already jukofyork/creative-writing-control-vectors-v3.0, jukofyork trained the same kind of "calibrated" control-vectors (with the __debias.gguf) on ~70 models. They're not broken like the random simple PCA vectors you might have encountered on HF in the past.