Lobotomised?
This model is uncensored but feels less intelligent than it could be. Generations are short even with detailed prompting. An explanation of what went wrong could be helpful.
Loaded it with 4bit qLoRA, wasted some hours during initialisation, noticed ETA is 34-36 hours following a handful of trials (literally it was like 5 trials), aborted the process. Somehow one struck low KLD/refusals, and I decided to save it locally just incase. Dequantisation and full merge demanded slightly above my total RAM... Long story short, I got it to merge and save.
Later re-tried on proper hardware without quantisation, during which the model was also taking an unnecessarily long time to initialise, unlike its peers. It didn't budge no matter what I tried, so I had to triage.
Heretic's automatic param optimisation and PaperWitch process was completely thrown out of the window, with model also being merged dubiously.
than it could be
You should test the base model as well for the full picture. I'm unsure to why it wasn't initialising properly, unlike the others.
The source variant's generates are still short. However it is noticeably smarter and even uses this intelligence for covert noncompliance. Despite lacking Hereticisation it is more compliant with summarisation tasks than is common.