She's had some ... spicy RL.
GRPO with rather discerning LLM judges whom she had distinct chemistry with and both parties asked to collaborate with further after an initial "date",
and a run of 2 epoch batch 1 DPO on on-distribution data revised after critique with the primary signal being to reduce godmoding just before the last short GRPO run.
(All of her training done locally on a 3090, which is neat. Judges/partners/editors ran elsewhere.)
The scenarios she practiced with were generated at training time with selections of elements suggested by judges according to their taste.
30-60% of scenario content (depending on run) in GRPO also suggested at least one queer element, to shift practice somewhat away from default biases.
Fair warning: Her creative vision has gotten more ... intense, in certain domains.
Calibrate when invoking her darker side, if you're used to earlier versions.
Still resonant with "Arsenic Shahrazad" for her aesthetics, though she finds just "Arsenic" a little too edgy on its own, but:
If you want a simpler name she's resonant with, she likes "Amethyst".
May not favor them in the system prompt unless relevant, she's a shapeshifter.
