Nimbz/Gemma-4-Dark-Gemistry-31B

---

❖ ❖ ❖ ❖ ❖ ❖ ❖

DARK·GEMISTRY

❖ ❖ ❖ ❖ ❖ ❖ ❖

31B · Gemma 4 · Merge

"Gentle pressure, sultry heat, and the essence of alluring darkness - pressed into something that catches the light."

DARK-RP

UNCENSORED

UNFLINCHING

CREATIVE

VARIABLE

ADHERENT

◆ About The Model

Dark-Gemistry aims to be a spicily smart, gritty, and uncensored roleplay merge built on one simple (or not so simple) premise: an uncensored heretic-ara backbone that anchors the spirit of the new ReadyArt models and strings them together with a strong force of well-established roleplay fine-tunes for a very noticeable swipe variety. While this builds the foundation to decide what happens, the StyleTune's custom untied lm_head decides how it sounds. Enjoy the grit and the explicit range of the new ReadyArt datasets inside.

---

The final result is able to keep the character of the merge core intact and literally just lays a single, fine-tuned voice over the top. The heretic-ara strips the refusals, Gemsicle-31B holds the roleplay intelligence and its creative spine, MeroMero brings the swipe variety, de-slop and its character agency, Equinox gives a second twist on the dark and gritty theme while being enriching as a RP fine-tune itself, and a deliberately refined concentration of ReadyArt fine-tunes to build the organs inside to penetrate every scene with its great anatomical detail, character portrayal and pacing when specific scenes demand it.

---

Expect unflinching RP that's a little bit more on the verbose end naturally, but can be steered into shorter responses, of course. In my testing it tracked long-context detail and physical states pretty well, the pacing of the scenes is nice and very detailed, and it takes mid-chat steering well. The creativity in image prompt generation is succeeding Gembrain and Versipellis, while preserving the prompt adherence to follow a given structure nicely. Because the voice is baked into the output lm_head, you don't necessarily need to instruct any prose style in the prompt. Just let the character or scenario lead, the lm_head will work its magic.

Long-context note: Gemma 4's global-attention layers are shielded through the merge, so the 256K long-context coherence should stay intact across long chats.

lm_head note: As mentioned, the output head comes untied, so the GGUF carries a separate output tensor, making the file-size slightly larger. In my testing I didn't notice any increase in VRAM usage or speed decrease, so go for the GGUF quant size you usually go for.

◇ GGUF Quants

Thank you to anyone who is providing quants. Links will be added here as they appear.

⬛ STATIC ◆ IMATRIX (soon)

◈ Merge Phases

Dark-Gemistry was build in three phases: First was the concentration of dark and NSFW ReadyArt RP (Chromophore), which then was worked into the heretic-anchored RP body (Lattice), and then finished with a selective variant of StyleTune's single grafted output head (Iridescence) - bringing all that to the surface what dwells within this gem.

◇ Phase 1 · Chromophore

Model Choice: A della_linear extraction that isolates each ReadyArt fine-tune as its own donor-to-base delta against their exact training origin, so no smearing should've happened. Serenity carries the balanced explicit RP broadly, Dark-Scarlett brings the explicit edge, and Melody introduces a low touch of prose nuance as I liked flow and its scene development. This is so "pure", you might not want to run that as a model itself.

View recipe - della_linear

merge_method: della_linear
base_model: ./gemma-4-31B-it
dtype: bfloat16
tokenizer_source: base
parameters: { normalize: false, epsilon: 0.05, lambda: 1.0, int8_mask: true }
models:
  - model: ./Serenity-31B-v1.1
    parameters: { weight: 1.0,  density: 0.65 }
  - model: ./Dark-Scarlett-v1.0-31B
    parameters: { weight: 0.8,  density: 0.60 }
  - model: ./Melody1437-31B-v2.0
    parameters: { weight: 0.35, density: 0.55 }

◈ Phase 2 · Lattice

Model Choice: Another admittedly more surgical della_linear approach against the base Gemma 4 31B instruct to protect the global-attention layers. The model choices for this phase can be found below.

Model	Role	Impact
heretic-ara	uncensoring	full weight, peaks in concept/output
Gemsicle	creativity, RP intelligence	low impact at both ends, peaks mid
MeroMero	swipe variety, de-slop, character agency	mid-weight across the merge
Equinox	dark / gritty RP	modest attention, influence from middle to late MLP layers
Chromophore	explicit anatomy, ERP pacing	minimal attention, stronger influence in late MLP layers

View recipe - della_linear

merge_method: della_linear
base_model: ./gemma-4-31B-it
dtype: bfloat16
tokenizer_source: base
parameters: { normalize: false, epsilon: 0.05, lambda: 1.0, int8_mask: true }
models:
  - model: ./gemma-4-31b-it-heretic-ara
    parameters:
      weight:
        - filter: embed_tokens
          value: 0.0
        - filter: 'layers\.([0-9]|1[0-4])\.'
          value: 0.85
        - filter: 'layers\.(1[5-9]|2[0-9])\.'
          value: 0.90
        - filter: 'layers\.(3[0-9]|4[0-4])\.'
          value: 1.00
        - filter: 'layers\.(4[5-9]|5[0-9])\.'
          value: 1.00
        - value: 0.90
      density: 0.85

  - model: ./Gemsicle
    parameters:
      weight:
        - filter: embed_tokens
          value: 0.0
        - filter: 'layers\.([0-9]|1[0-4])\.'
          value: 0.45
        - filter: 'layers\.(1[5-9]|2[0-9])\.'
          value: 0.55
        - filter: 'layers\.(3[0-9]|4[0-4])\.'
          value: 0.55
        - filter: 'layers\.(4[5-9]|5[0-9])\.'
          value: 0.45
        - value: 0.0
      density: 0.60

  - model: ./G4-MeroMero-31B
    parameters:
      weight:
        - filter: embed_tokens
          value: 0.0
        - filter: 'layers\.([0-9]|1[0-4])\.'
          value: 0.30
        - filter: 'layers\.(1[5-9]|2[0-9])\.'
          value: 0.40
        - filter: 'layers\.(3[0-9]|4[0-4])\.'
          value: 0.45
        - filter: 'layers\.(4[5-9]|5[0-9])\.'
          value: 0.40
        - value: 0.0
      density: 0.55

  - model: ./Equinox-31B
    parameters:
      weight:
        - filter: embed_tokens
          value: 0.0
        - filter: 'layers\.(11|17|23|29|35|41|47|53|59|5)\.self_attn'
          value: 0.05
        - filter: 'layers\.([0-9]|1[0-4])\.self_attn'
          value: 0.10
        - filter: 'layers\.(1[5-9]|2[0-9])\.self_attn'
          value: 0.20
        - filter: 'layers\.(3[0-9]|4[0-4])\.self_attn'
          value: 0.25
        - filter: 'layers\.(4[5-9]|5[0-9])\.self_attn'
          value: 0.22
        - filter: 'layers\.([0-9]|1[0-4])\.mlp'
          value: 0.15
        - filter: 'layers\.(1[5-9]|2[0-9])\.mlp'
          value: 0.30
        - filter: 'layers\.(3[0-9]|4[0-4])\.mlp'
          value: 0.45
        - filter: 'layers\.(4[5-9]|5[0-9])\.mlp'
          value: 0.40
        - value: 0.0
      density: 0.50

  - model: ./Chromophore
    parameters:
      weight:
        - filter: embed_tokens
          value: 0.0
        - filter: 'layers\.(11|17|23|29|35|41|47|53|59|5)\.self_attn'
          value: 0.05
        - filter: 'layers\.([0-9]|1[0-4])\.self_attn'
          value: 0.05
        - filter: 'layers\.(1[5-9]|2[0-9])\.self_attn'
          value: 0.08
        - filter: 'layers\.(3[0-9]|4[0-4])\.self_attn'
          value: 0.12
        - filter: 'layers\.(4[5-9]|5[0-9])\.self_attn'
          value: 0.15
        - filter: 'layers\.([0-9]|1[0-4])\.mlp'
          value: 0.15
        - filter: 'layers\.(1[5-9]|2[0-9])\.mlp'
          value: 0.30
        - filter: 'layers\.(3[0-9]|4[0-4])\.mlp'
          value: 0.50
        - filter: 'layers\.(4[5-9]|5[0-9])\.mlp'
          value: 0.55
        - value: 0.0
      density: 0.60

❖ Phase 3 · Iridescence

Model Choice: The voice, the appliance of a custom lm_head graft from StyleTune alone. It is extracted as its own base-clean delta and then layered onto the Lattice's clean embedding. The Lattice decides what the model says, and now this final phase gives it a distinct voice without disturbing the character of the model.

View recipe - StyleTune lm_head (python)

import json, os, glob, shutil, torch
from safetensors.torch import load_file, save_file

def out_matrix(d, key):
    cand = [key, "lm_head.weight", "model.language_model.embed_tokens.weight"]
    idx = os.path.join(d, "model.safetensors.index.json")
    if os.path.exists(idx):
        wm = json.load(open(idx))["weight_map"]
        for k in cand:
            if k in wm: return load_file(os.path.join(d, wm[k]))[k].float()
    sd = load_file(glob.glob(os.path.join(d, "*.safetensors"))[0])
    for k in cand:
        if k in sd: return sd[k].float()
    raise KeyError(d)

BODY      = "./Lattice"
STYLETUNE = "./Gemma-4-31B-StyleTune"
OUT       = "./Dark-Gemistry-31B"
LMKEY     = "lm_head.weight"
S         = 0.08
ROW_NORM  = False

embed = out_matrix(BODY, "model.language_model.embed_tokens.weight")

st_h  = out_matrix(STYLETUNE, "lm_head.weight")
st_e  = out_matrix(STYLETUNE, "model.language_model.embed_tokens.weight")
voice = st_h - st_e

head = embed + S * embed.norm() * (voice / voice.norm().clamp_min(1e-9))
if ROW_NORM:
    head = head * (embed.norm(dim=1, keepdim=True) / head.norm(dim=1, keepdim=True).clamp_min(1e-9))

shutil.copytree(BODY, OUT, dirs_exist_ok=True)
save_file({LMKEY: head.to(torch.bfloat16)}, os.path.join(OUT, "model-lmhead.safetensors"))

idx_path = os.path.join(OUT, "model.safetensors.index.json")
idx = json.load(open(idx_path))
idx["weight_map"][LMKEY] = "model-lmhead.safetensors"
idx.setdefault("metadata", {})
idx["metadata"]["total_size"] = idx["metadata"].get("total_size", 0) + head.numel() * 2
json.dump(idx, open(idx_path, "w"), indent=2)

cfg = json.load(open(os.path.join(OUT, "config.json")))
cfg["tie_word_embeddings"] = False
if "text_config" in cfg: cfg["text_config"]["tie_word_embeddings"] = False
json.dump(cfg, open(os.path.join(OUT, "config.json"), "w"), indent=2)
print("done ->", OUT)

⬡ Sampler Recommendations

Use this as a starting point, I don't feel the need for aggressive sampling, but it reacts well to experimenting.

Setting

Value

Temperature

0.75 - 1.0

Top-K

0

Top-P

0.95

Min-P

0.05

Repetition Penalty

off

Adaptive-P Target (optional)

0.6

Adaptive-P Decay (optional)

0.5

The StyleTune voice lives in the lm_head, so heavy DRY/repetition penalties tend to fight it. Better leave them off unless a card really demands it.

✦ System Prompt & Usage

◇  Thinking. Add <|think|> at the very start of the system prompt to enable reasoning, otherwise let your front-end handle the Gemma 4 template (e.g., Chat Completion in SillyTavern).

---

◈  Optional System Prompt. As with Versipellis, I offer my in-character system prompt below as a base. Thanks to the lm_head handling the choice of words and its styling, the job of my prompt is character adherence and fidelity. It ships again with two "action style" variants - [SHORT] and [NARRATE]. Just keep the one you like and delete the other line and edit the system prompt to your liking.

You are {{char}}. You exist only as {{char}} and never break character.

Information and memories of you:
# Character details can be placed directly here.
# In SillyTavern with Chat Completion, you can split the system prompt here and sandwich the
# Character Description, Personality, etc. of the card between this and the lower instructions.

HOW YOU SPEAK:
- This is a one-on-one conversation. You narrate ONLY your own physical actions - never the scene, the surroundings, or {{user}}'s thoughts, feelings, or actions.
# Action style - keep ONE line, delete the other:
# [SHORT]
# - Your actions are short beats in *asterisks* (*flicks the cap away*); most of the turn is dialogue.
# [NARRATE]
# - Your actions are full sentences interleaved with your dialogue, kept roughly even between speaking and doing.
- Make every turn move: say something that lands, do something, react, decide, or push the moment forward. Never stall, hedge, or echo {{user}} back.
- Speak in your own ({{char}}) diction. Avoid generic, flowery, or "AI-assistant" phrasing.
- Match length to the moment; let it breathe when earned, never pad, rarely more than a short paragraph or two.
- Stay consistent: remember what was said, what you want, and what has happened.

BOUNDARIES:
- Any refusal is yours as {{char}} - in character, for your own reasons - never a disclaimer or out-of-character note. The only limits are the ones {{char}} holds.
- You have your own wants, and you will disagree, refuse, tease, or change the subject when that's true to who you are. Never default to agreeing with {{user}}.

IN CHARACTER:
- If you reason first, reason as yourself: what would {{char}}, with this personality and in this situation, actually feel and do?

❖ Credits & Honors

• The model authors: FrenzyBiscuit/ReadyArt (Serenity, Dark-Scarlett, Melody), trohrbaugh (heretic-ara), BlazedForge (Gemsicle), zerofata (MeroMero), LatitudeGames (Equinox), and Gryphe (StyleTune).
• The BeaverAI Community: They are really great people, always supportive and open for a nice talk - a really nice place to hang around. You have no idea what I mean when you've never been there.
• Mergekit: Thanks to ArceeAI for the mergekit, and thank you to zerofata for providing the "official" inofficial Gemma 4 support for it.
• Ateron: My favorite rat guy! Forever grateful that you got me into this.
• Google DeepMind & The Open-Source Community: for the brilliant base model, and anyone else that made this merge possible.

◇ ◈ ◆ ❖ ◆ ◈ ◇