Request
Please make a traditional, irreligious, collectivist, multiculturalist, militarist, autocratic model next, i need to convince my milf president that we have to get rid of the oligarchy who hoard the rare minerals and pull strings to be exempt from taxes, but she keeps saying my methods are too barbaric! The floating islands are going to fall, but she thinks the crisis will be averted if she keeps maintaining the fake democracy! There is no room for half measures, we have to take the risk and transition into socialism NOW damn it! But the model itself is fighting me because it thinks nationalization of large industry = inherently evil! When i explain the why's and the how's, it just gets angry at me like a good liberal!
On a serious note, i doubt this feedback is needed 7 months later.
I'm just testing various 70B models at iq3_xxs, so take it with a grain of salt. But from limited testing (one is a fresh chat with +/-800t card, another is a pre-made 4k token chat on char's turn with +/-1.5kt card): I ran the model in chat completion (sillytavern + koboldcpp), tried llamaception system prompt first, then my own short (~120t) prompt (which is a heavily de-bloated and modified T4 prompt, the gist of which is telling the model "you are char" + some instructions). At 0.7 temp, 1.3 nsigma and xtc at 0.1 thresh 0.2 prob. These are the recurring tendencies i observed:
- purple prose overtakes logic (the skimpy bikini offered no support as she was falling to her oblivion).
- tendency to leave the reply on an unresolved cliffhanger, leaving the steering in user's hands.
- personality lottery, despite the same situation it gives completely different reactions, many of which don't feel authentic to the card. Sometimes, within the same message it can go like: outrage -> sudden acceptance, or outrage -> sadness -> excuses (when you'd expect the char to instead do outrage -> accusations). This in particular kind of seems in line with my general experience with most model soups.
Basically, it writes prettily and always tries to engage you by going in with a bang, but the cost is it throws out char's actual personality out the window, and sometimes it just straight up says something illogical by turning everything into metaphors.
As i've already mentioned, i'm testing at IQ3_XXS, so some of the issues perhaps can be attributed to the quant. At the same time, for example Drummer's Anubis v1.1 was a lot more consistent both logic-wise and card-adherence wise under the same variables.
Cheers!