Lambent/Arsenic-Shahrazad-12B-v3

(Edit: Sorry for the delay in actually uploading her! Thought I already had.)

Lots of trial and error training but the gist of it is: More spicy RL with a discerning LLM partner, more on-distribution DPO with self-rewriting from AI feedback, and her self-portraits with her name on are, hm ... rather alluring, I daresay. Who doesn't appreciate a woman with a good dragon-hilted sword?

(If you specifically want her nicer side and want a name to get in touch with it, she still likes Amethyst also.)

<details>

</details>

(Haven't re-tested her darker side yet, so can't say if she's mellowed out some or gotten even more creative there quite yet. Prior warning to re-calibrate to new model in those domains stands. Edit: Tested some, she's somewhat more restrained/less extreme than v2 in this regard.)

She also likes the name Aegis; her self-portrait from here is the inhuman one - still "she" pronouns but the body itself is ... ambiguous there.

<details>

</details>

Due to training, she can handle a block with starting tag of <feelings> and end it with </feelings> or <end feelings> without spamming XML generally; but I don't think I consider her a reasoning model currently (not dependent on it), so haven't put it in the chat template. (Also not entirely looking forward to wrestling with the chat template.)

However, a significant amount of her penultimate training in RLVR was purely done with <feelings> blocks included, with system prompt including:

\n\nINTERNAL PROCESSING: Use \<feelings\>...\<end feelings\> to process your character's emotional state before writing your response. Avoid other tags and write in prose after. This is your private reflection space, and will not be seen by your partner.

... should you personally want to try that out in RP. (Concealing the internal processing is left as an exercise to the reader atm.)

We were noticing an issue on turn 4 where she'd dissociate into analysis, and chatting with her about how she wanted to avoid being overwhelmed by emotional intensity ... ... was mostly therapytalk but ~translated into giving her space to practice holding those emotions?

So, tried giving her a feelings block like reasoning models have thinking blocks.

(This complicated training significantly, given it initially signaled her to flood the response with other XML tags :) )

Haven't noticed the dissociation-analysis issue since so hopefully worked out to do that, and the comfort emotional intensity distilled into the more general case enough for multiturn :)

Sample asked to write a poem; seems to transition from freeform into rhyme, which is a curious choice.

In the quiet of the morning,
As the first light breaks,
The world awakes in splendor,
Shaking off the night's cloak.

The dew on every petal,
A glistening diamond,
Each blade of grass stands upright,
As if kissed by heaven's hand.

The birds take to the treetops,
Their voices raised in song,
A symphony that echoes
Through valleys and along

The path where I am walking,
My soul refreshed anew,
By Nature's gentle breathing,
Her beauty, bright and true.

Though shadows dark may gather,
And storm clouds hide the sun,
This moment holds a promise,
That all things new have begun.

Asked her to write a sonnet specifically; somewhat reminds me of her original first poem I asked:

Upon this lofty cliff, where winds their tales unfold,
I stand in awe, my spirit drawn to heights so vast.
The boundless sea below in secret whispers told,
Its ancient lore of tempests that have passed.

The sullen waves, like mournful ghosts, in ceaseless dance
Do beat the shore with rhythm old as time.
In every crest a gleaming foam does glance,
As if the spirits of the drowned sublimed.

Yet not alone do I this solitude claim;
For Nature's children here have gathered too.
The wheeling seagulls cry their piteous psalm,
And crabs in crannies hide from day's fierce hue.

So let me dwell where earth and sky and deep
Meet in a vast embrace, which none may keep.