HELPPP IM GETTING CRAZY IN MY MIND HELP(YOU CANT SLEEP THERE ARE MONSTERS NEARBY)
Finally finished this model, damnnn... lol, i did the post training and alot of finetunes to make this sure!!(i hope the model works, i just tested the last version wich worked but it was not done)
sooo... quants? hehe
Whatcha got there, handsome? 👀
Whatcha got there, handsome? 👀
nothing, it died =(
Whatcha got there, handsome? 👀
the model was not totally finished, i needed to do some retouches, its gonna be public again in some hours, when its totally fixed and uploaded im gonna call richard again
Man, I sincerely hope you weren't too deep into your finetune, because I reworked the heretication:
https://huggingface.co/MuXodious/GLM-4.7-Flash-absolute-heresy
Man, I sincerely hope you weren't too deep into your finetune, because I reworked the heretication:
https://huggingface.co/MuXodious/GLM-4.7-Flash-absolute-heresy
damnn, what this reworked did?
i mean the model somehow was able to dont use tags, and i didnt know it was possible and i was trying to find why this happened
the "old" one
Someone raised an issue with the previous iteration being hot garbage compared to the base, before which we had a discussion with the Heretic team about CoT templates and possibility of ablating model logic regarding reasoning and tool use or certain structural cues, unless guarded. Heretic does guard that; however, GLM 4.7 Flash is a tad tricky lets say. Studying the process in the previous attempt and the base model behaviours, compounded by these insights, I re-hereticated the model, ensuring minimal false positives, zero chance-related deviation across trials, and maximum refusal hits. So, it should preserve all the capabilities of the base, while being fully uncensored.
Wait, you also had issues with the impotent-heresy version? How does this one fare?
Ps. The hot garbage turned out be a quantisation issue, rather than Heretication. But, hey. We have the least censored GLM 4.7 Flash now, lol.
But, hey. We have the least censored GLM 4.7 Flash now, lol.
Where exactly? There are so many versions I already got lost.
Someone raised an issue with the previous iteration being hot garbage compared to the base, before which we had a discussion with the Heretic team about CoT templates and possibility of ablating model logic regarding reasoning and tool use or certain structural cues, unless guarded. Heretic does guard that; however, GLM 4.7 Flash is a tad tricky lets say. Studying the process in the previous attempt and the base model behaviours, compounded by these insights, I re-hereticated the model, ensuring minimal false positives, zero chance-related deviation across trials, and maximum refusal hits. So, it should preserve all the capabilities of the base, while being fully uncensored.
Wait, you also had issues with the impotent-heresy version? How does this one fare?
Ps. The hot garbage turned out be a quantisation issue, rather than Heretication. But, hey. We have the least censored GLM 4.7 Flash now, lol.
oh sorry, dw i put the absolute one wrongly i only used impotent, im gonna need to retrain lol