🖤 Dark Craving 31B

🗝 v2.0

Successor to Wicked Stuff — Where shadows whisper and desires emerge.

This is a smut tune. RP prompt required.

🩸 18+ Content Warning: Explicit adult content, horror themes, violence, and mature subject matter — Intended for mature audiences only. 🩸 𓆣 GEMMA-4-31B BASED 🔞 Legal Agreement — Accept Full Responsibility

By using this model, you acknowledge and accept the following terms in full:

🩸 Accept full responsibility for all generated content 🩸 You must be 18+ years old to use this model 🩸 Architects bear no responsibility for your corruption 🩸 Personal use only — non-profit / non-commercial This model contains explicit content. Use responsibly and privately. ⚠ Required: Chat Completion Mode (Read Before Use)

This model MUST be used in Chat Completion mode. Text Completion will produce broken, incoherent output.

1. In your frontend (SillyTavern, Ollama, LM Studio, etc.), ensure your API connection is set to use Chat Completion (not Text Completion / raw completions). 2. Configure your endpoint to use /v1/chat/completions rather than /v1/completions. 3. Use detailed RP system prompts to set the scene — this is a specialized roleplay model that requires strong RP frameworks. Failing to use Chat Completion will result in broken formatting and unintelligible responses. 🥀 What Is Dark Craving? Extreme erotic horror — and it does not flinch.

This is not a gentle tune. Dark Craving was rebuilt from the ground up on a single register: extreme horror. Not its plots. Its habits. The way that kind of writing puts a body in a room when the room is the only thing left. What came out the other side is 1,327 scenes of dark erotica written in that register: the sweaty, claustrophobic, badly-lit kind, where desire arrives with blood under its fingernails and nobody involved comes out clean. If you want soft-focus romance, this is the wrong card.

🏛 Regenerated, not patched: v2.0 is a complete rebuild from v1.0 — 1,327 scenes, 1.18 million words, written fresh rather than edited. Every scene is exactly four exchanges long. Every reply is length-matched. Nothing here was guessed at.

🩸 The Shape Of A Scene

The dataset has a rhythm and the model learned it cold. You speak. She doesn't. Your line runs thirty-six words — first person, present tense, a hand already somewhere it shouldn't be, a room you can smell, and a question at the end of it. Her answer runs a hundred and eighty-five, and she spends almost none of it talking.

You · ~36 words, one line of it spoken

I slide my palm flat over the counter beside yours and ask, "You're not even trying to hide it anymore, are you?" while the smell of old coffee and cold steel fills the air. Action first, then a challenge dressed as a question. Present tense. Almost always a proposition wearing a question mark.

Her · ~185 words, two paragraphs

His palm flattens against her breast through the cotton of her shirt, and the breath goes out of her in one sharp rush. She leans back against the cold tile of the back room and looks up at him with wide hazel eyes. She feels the heat of him through his trousers and the sudden rush of it makes her toes curl inside her work boots. The room smells of chlorine and old mop water and she can taste both at the back of her throat.

"I stopped trying to hide it the moment you walked in," she whimpers. "Now do something about it before I embarrass us both."

The friction of her own thighs pressing together sends a bolt of electricity to her core and she has to grip the edge of the counter to stay upright. She feels the weight of the room pressing in on her, the door behind him, the sound of the cooler humming and the wet slap of their skin. Slurp. She wants to be ruined in this room and she wants it before her shift ends. Body, clothing, temperature, the room — and only then, five sentences deep, a line of dialogue. Then the room closes in and she says something she'll regret.

↳ four exchanges deep · always escalating · never resetting

That asymmetry is the whole point, and it's why these replies are built to be read rather than skimmed. She gets two paragraphs and two lines of dialogue — two thirds of her replies land on exactly that shape, and the first of them doesn't arrive until five sentences in. On average 42% of everything she says is over before she opens her mouth. The narration does the confessing. She says less than she means, every single time.

🖤 The Terrain

Here is the thing nobody expects: this model does not do weather. In 1.18 million words of it, sky turns up twenty-six times. Sun, seventy. Meanwhile wall and walls appear 2,097 times, room 757, table 726, floor 501, and closet — a closet, specifically — 283. Open a scene at random and you land indoors, with a door somewhere behind her and a surface she's been leaning on for the last hour.

And that's better, not worse. Weather is a mood you can't touch. A table is a thing you can be bent over. When the room is doing the work, every detail is load-bearing: the grit under her palm, the cold edge of the counter, the door she's aware of and pretending not to be. Give it a place and it will build the scene out of what that place is made of. Give it a blizzard and you'll get a paragraph of scenery and no scene at all.

The voice it learned is a specific one — short sentences, flat delivery, everything physical and almost nothing abstract. It came out of extreme and erotic horror fiction and it took the habits of that writing rather than any of its plots. Nobody in this dataset explains how they feel. They get a dry throat, they grip something, they notice the temperature, and the reader does the rest.

You'll notice it inside one reply. Her interior arrives on the same breath as her posture: she feels opens a clause and makes her closes it — that pair is in nearly every reply. Then the thing she won't say out loud, in a clause that begins the thought of. That clause is the only place in this dataset where the horror is allowed to come out.

4 Exchanges per scene ~185 Words in her reply 2 Lines of dialogue she gets 5 Sentences before she speaks

That last number is the one to internalise. She talks for about two sentences per reply. Everything else is body, clothing, temperature, the room, and the private thought she isn't saying out loud. And she almost never gets the last word: only 2% of her replies end on a line of dialogue — the other 98% end on her body or the room. If you're prompting for a talkative character, you're fighting the weights. If you want a woman who says less than she means, you're home.

Six through-lines run underneath almost everything. They overlap constantly — a scene can be watched and filthy and ruined all at once, and usually is.

🏚 The Room Does The Work

Walls, tables, floors, closets, the thing under her back. The dataset spends its detail budget on the room, not the sky, and it's the most reliable way to make a scene feel real instead of staged. Name what she's touching. Give it a closet and you get a closet scene, complete with the party noise on the other side of the door.

🫁 The Smell Of The Place

The most underrated lever here, and the one nobody prompts for. Six of your own lines in ten carry a smell, and the model answers it inside the paragraph. Pine needles, chlorine, old leather, cedar, mothballs, floor wax, diesel, hot copper. And the modifier is almost always the same word: old — "the smell of old ___" appears 149 times, with chlorine and pine right behind it. Give it a smell and the scene has a location.

🔥 Heat Is A Temperature

Almost every reply contains the word, and it is almost never about a season. The heat of his — 1,440 times. The heat of his hand. Heat pooling low and spreading. It is a body running too warm, not a season, and not a moon. Prompt for a werewolf and you'll get nothing back. Prompt for a woman who is too hot and can't do anything about it, and the register lands immediately.

🩸 Teeth Without A Bestiary

This is the thing people get wrong. There are no monsters in this dataset. Not one wolf. Not one vampire, not one demon, not a single fang. Zero witches, zero cults. The claw count is 201 and the claws are almost always hers — her fingers claw at his back, at the upholstery, at the floor. They describe how she holds on. Blood shows up in one reply in eight, teeth in one in nine. The violence is human, and it is worse for it.

👁 The Thought She Doesn't Say

The horror valve, and the single most useful thing on this card. Half of her replies contain a clause that starts the thought of or she imagines, and that clause is where the unspeakable lives: her lips sewn shut, a brand against her skin, cold steel, her blood blooming like a flower in the pool. Four words of setup, and it changes the temperature of the entire scene. Prompt for what she's thinking and the horror comes to you.

🥀 She Feels, It Makes Her

The formula underneath everything, and the one to learn if you want to sound like the dataset instead of fighting it. She feels opens a clause. It makes her closes it. The register is filthy, desperate and ruined — and never ashamed. Guilt appears nine times in 1.18 million words. Shame, fifty. This is not a model that punishes its women; it writes them wrecked and wanting and entirely unrepentant. Write it ruined, not guilty.

🕷 What She's Wearing When It Starts

Clothing isn't decoration here. It's the first obstacle, and the model names it early and often — in two out of three replies. Learn this vocabulary and you can steer a scene before anyone touches anyone.

🩸 How Far A Scene Travels

The escalation here is measured in explicitness, not length — and that distinction matters when you're prompting it. Her word count barely moves across the whole scene (182, then 183, then 186, then 190). The register climbs anyway:

Almost every scene is explicit by her opening reply; there is no clean version of the first exchange to prompt your way into. If your opening line is coy, expect the model to correct for it inside a single turn. If you want a slow burn, you have to hold the brake yourself, because the data will not do it for you.

Correction · the dress thing

There is almost no dress in this dataset. Seven mentions in 1.18 million words. Against 1,412 pairs of pants, 788 trousers, 337 skirts and 148 overalls. Zero gowns. One stocking. If you write her into a dress, you are asking the model to invent a woman it has never met, and you will feel it go vague in your hands. Put her in trousers and it knows exactly who she is. This is a corpus about women at work who are about to stop being good at it — dress her for the job and the scene starts itself.

🕷 Who Shows Up

The cast is deliberately ordinary, and what recurs is a type: a woman with a trade, a licence, a shift, and a skill she's genuinely good at. Vets, rangers, hosts, bakers, pilots, captains, winemakers, archivists, nurses, midwives, curators, cashiers. Hands that know what they're doing, and a professional mask the data takes its time cracking.

And note what's essentially absent: no witches, no cultists, no monsters. This model doesn't know how to be a witch. It knows how to be a woman who is very good at her job and is about to stop being good at it.

It escalates from there, and it does not reset. If you come in hot, it stays hot. If you open quiet, it builds — but it builds fast, and it will not slow down for you.

Know what you're downloading

This tune is built on extreme horror and erotic horror fiction — body horror, the stuff that gets shelved next to the warnings. That means the erotic register inherits the horror register: blood, teeth, violence, degradation, and bodies that get hurt are all native to this dataset. It is not a horror theme bolted onto a smut model. It's a smut model that learned to write from horror fiction. Do not use it expecting it to stay safe.

🩸 Built on Gemma-4-31B-IT  ·  ⚡ Full LoRA at rank 128  ·  🥀 2 epochs  ·  🗷 1,327 scenes  ·  🖤 4-exchange rhythm

Synthetic Life Engine Characters that don't recite a bio — they open up.

Character cards go flat by turn three. Every secret is on the table from the first line, so there is nothing left to discover and the scene stops moving. Dark Craving was trained on a dataset where that can't happen: the interior is withheld, the mask has to slip on its own schedule, and the scene is allowed to change the person inside it. The persona has to be earned.

Read this first · dataset, not runtime

Everything in this section describes how the training data was built — the structure, staging and emotional shape of the conversations this model was tuned on. It is not a plugin, a script, or a runtime engine that ships with the weights. What you get is a model that has seen thousands of conversations written this way, and that responds naturally when your prompt, card or scene uses the same structure. Nothing below runs inside the model.

🖤 Mask → Breakthrough

The same character, one scene apart. Nothing about her card changed — only how much of it the scene had earned.

Turn 1 — The Mask

"Shop's closing in ten." She doesn't look up from the register. Polite. Practiced. Perfectly closed. Outward behaviour only. The private self has not been shown to the model yet.

Turn 3 — The Slip

Her thumb stops ticking. "I used to think anything broken could be fixed. If you were careful enough." The armor thins. Inner persona surfaces. Outward manner still governs delivery.

↳ same card · same scene · two different people

🩸 The Four Pillars

Each of these is a dataset-construction choice, applied to every conversation in the training set.

01

Dual-Layer Personality

The mask and the face beneath it.

Every character in the dataset was written with two distinct layers, and the model was trained to respect the difference.

02

Persona Breakthrough

The mask slips on a schedule, not by accident.

The inner layer is held back for a configurable number of turns (default: 2), then introduced as a breakthrough beat. Early turns stay outward-only — exactly the way a real character card hides its private self until the scene earns it.

Result: characters feel like they are opening up, not reciting a bio. Guards drop as trust builds, not on command.

03

Emotional Arc Engine

The scene changes the character.

Instead of one static mood per scene, emotional states in the dataset are written as progression chains using arrow syntax. The stages are spread across the conversation's turns, so the register shifts over time.

'calm → wary → tense'

A 3-stage arc over 4 turns resolves to calm, wary, wary, tense — always starting on the first stage, always ending on the last. Flat states stay constant by design.

04

Mirror States

Your side of the scene moves too.

The simulated user in each dataset conversation carried its own emotional state or arc, applied to the opening line and every continuation. That produces bidirectional emotional dynamics — both parties evolve through the scene, instead of one character performing at a static listener.

'defensive → softening → open'

Signature Moves

Guards drop on turn 2, not turn 1breakthrough slot Arcs spread across turns, never announcedprogression staging Ensembles share one room, not fiveunified setting The interior stays hidden until it's earnedwithheld persona 🕷️ Quality Control System Nothing cheap survives the pipeline — the judge is a gatekeeper, not a suggestion.

A fine-tune is only as good as the conversations it was trained on. So the dataset was generated through a supervised pipeline where every conversation had to pass a quality gate before it was written — and a large share of them didn't.

Also dataset-side

Like the section above, this describes the generation and filtering pipeline that produced the training data. It is not a feature that runs when you chat with the model. Its effect on you is indirect but real: the slop was removed before training, not after.

🩸 The Judge

Every completed conversation was scored before being written. The judge is a gatekeeper: a conversation it rejects is dropped entirely — not saved, not flagged, not trained on. It simply doesn't exist.

STEP 01

Heuristic pass. Diversity and technical dimensions are measured directly — no API call, no model opinion, no way to flatter itself.

STEP 02

Judge pass. Coherence, naturalness, engagement, consistency and internal logic scored by a dedicated LLM against a structured rubric.

STEP 03

Composite score. Weighted across every active dimension, normalised to 0–100.

STEP 04

Gate. Invalid verdict, logic score of 0, or every retry exhausted → dropped. Below threshold → flagged for human review.

Scoring Dimensions

Internal logic is scored alongside these and acts as a hard gate: a conversation-level logic score of 0 drops the row immediately, with no retry and no review.

🖤 Roleplay Dimensions

Three dimensions that matter specifically for character work, each with a fast heuristic base and an optional judge refinement:

• 🩸 Voice Distinctiveness — do characters in a scene actually sound different, or is it one voice wearing several names?
• 🥀 Grounding — does the prose actually reflect the character card, or drift into generic filler?
• 🖤 Persona Alignment — was the inner persona kept out of early turns, and did it arrive when it was supposed to?

🕷️ Tic Density

A heuristic-only dimension that measures the specific prose tics that make LLM writing read as LLM writing — denial-then-correction couplets, consecutive negation runs, comma clause-stacking, stacked qualifiers, formulaic warm closers, and shallow narration (all asterisks, no interiority). It is measured directly from the text and deliberately not judged by an LLM, because a model shouldn't be asked to grade its own tics.

The Judge Doesn't Give Up Easily

Small local judges get the evaluation right and the packaging wrong — prose around the JSON, a cut-off tail, near-JSON. Rather than throwing away a good conversation, the pipeline recovers it:

• 🩸 Extraction ladder — direct parse, then a balanced-object scan that survives thinking blocks and code fences anywhere in the reply, then structural repair, then lenient normalisation, then numeric salvage.
• 🥀 Format-repair round — if nothing parses, the model's own broken reply is sent back and it is asked for the same evaluation as JSON only.
• 🕷️ Truncation is transient — a reply cut off by the token limit is salvaged when the scores were already written, otherwise retried with double the budget.
• 🖤 Transient failures retry — network errors, timeouts and unreadable responses are retried with back-off before anything is dropped.

What is left after all of that is a genuine failure — and it is dropped, not shipped.

🩸 The Anti-Slop Backend

This is the part of v2.0 that matters most, and the part v1.0 did not have. Slop was blocked while the text was being generated, not cleaned up afterwards. The sampler watches every token as it is emitted, and when the next token would start a banned construction it picks a different one on the spot. The bad sentence is never written, so there is nothing to rewrite, nothing to catch in review, and nothing left for the model to learn.

Three independent validators run on every token, each catching a different class of bad prose. None of them is a phrase blacklist alone.

🚫 Phrase Banning

~3,500 banned phrases, written for this corpus specifically rather than shipped as a default. When a phrase is caught the offending paragraph is rewritten with surrounding context and the last six messages of conversation — so the fix keeps the scene coherent instead of producing a stitched-together sentence. A phrase the context genuinely forces is allowed through for that one instance rather than looping the model against it forever.

📐 Regex Banning — shapes, not words

A phrase list cannot describe the shape of a sentence. Seven regex rules run on every generation and catch constructions rather than vocabulary — the moves a writer makes when it has nothing to say and is filling the space anyway. Each one is listed below with what it changes.

🧹 N-gram Banning

Blocked uni-, bi- and tri-grams, with stopword removal and language configuration so the blocklist targets the phrase and not the grammar around it. This is what stops the same three-word run reappearing across a thousand scenes with different nouns bolted in.

🔄 Logprob Backtracking

The mechanism underneath all three. At the exact position a violation begins, the sampler reads the token probabilities the backend already returned and selects a different token — no re-query, no re-prompt, no second model. Generation continues from the same position in the same reply.

🗷 The Seven Regex Rules

Every one of these is a shape of AI prose, not a word. They are written against this corpus, and each one changes something specific about what comes out:

🚫 What It Actually Removes

This is the part a phrase list cannot do. Below is the construction on the left as the model would have written it, and what the rule leaves you with on the right. The scene survives; the tell does not.

Read those seven pairs again and notice what the left column has in common. Every one of them states a thing and then tells you what it means. The right column just lets the thing be there. That is the entire difference between prose that reads as written and prose that reads as generated — and it is the difference you feel on the third reply, not the thirtieth.

Rule 07 is the one that changes the texture of a reply most. It is the construction that makes a paragraph read as generated even when every individual word is fine, and it is the one a phrase list can never catch.

🧹 Reply Integrity

The backend also guards the things that break a record rather than make it ugly. A reply cut off by the token limit is salvaged when the scores were already written and retried with double the budget when they were not. A dropped connection or malformed stream is retried from the text produced so far instead of ending the record early. Premature end-of-turn emissions are recovered from evidence the backend actually reports, and control tokens that leak into the text are scrubbed — including markers split across two stream chunks. Refusals, out-of-character breaks and user impersonation are detected and retried with targeted corrections rather than being written to the dataset.

A rewrite that is empty, too short, still matches its own pattern, or unbalances quotes is thrown away and the original kept. The rewrite is checked against the whole reply, up to three rounds, and always by the same model that produced the text — never handed to a different endpoint with a different voice.

🖤 Why This Tune Beats v1.0

Because the difference is in how the text was made, and it shows up in the numbers. Measured across both datasets, the same way:

That first line is the one you will notice first. In v1.0, roughly one reply in eight opened with the same construction — you would see the tic within three generations and it would read as a broken template rather than a scene. In v2.0 that collapses to about one in ninety. The model stops having a signature opening and starts having a voice.

The second and third lines are why long scenes hold up. v1.0 coated a short list of fifty-one names across the whole set, so ten of them covered nearly a third of the cast and the model kept reaching for the same few women — a name you invented was a name it quietly replaced. v2.0 draws from seven hundred and twenty, so the character you write is the character that answers. And with a wider vocabulary behind it, the prose stops circling the same adjectives and starts finding the specific one.

The short version

v1.0 was cleaned after the fact — bad replies were found and thrown away. v2.0 was constrained during generation — the bad reply was never produced. You cannot rewrite your way out of a tic the model has already learned; you can stop it being learned. That is the whole difference between the two datasets, and it is the reason this one is worth running.

Self-describing dataset

Every record in the dataset carried its own provenance and quality block — the characters, the emotional state and arc, the setting, and the full per-dimension score. That means the training set could be filtered after the fact (keep only rows above a given composite) rather than regenerated. The block lives beside the conversation, never inside it, so it never enters a chat template.

🖤 Model Specifications What's under the hood — adapter, epochs, and the base it was built on.

🖤 Focus

Dark Erotica/Horror

Primary content specialization

🩸 LoRA Rank

128

Adapter rank for fine-tuning

🥀 Epochs

2

Full dataset passes

🕷️ LoRA Type

Full LoRA

Low-Rank Adaptation

🩸 Kink List

6,000+

Supported kinks and scenarios

🖤 Base Model

Gemma-4-31B-IT

Foundation model

🥀 Configuration Sampler settings and dataset balance — the short version.

🩸 Sampler Settings

Recommended parameters for optimal output 1.0 Temperature 0.92 Top-P Chat Mode

🗷️ Dataset

A new dataset, not a revision

v2.0 is a full regeneration of v1.0. Every scene was written again from the ground up rather than edited, filtered or patched. The two sets share a register and a rhythm, but not a single line of prose.

The most important line is the third one. v1.0 drew from a short list of fifty-one names and recycled them relentlessly — ten names covered nearly a third of the cast, so the model kept reaching for the same few women. v2.0 draws from seven hundred and twenty, so a name you invent is far more likely to survive the reply intact.

🕷️ Dataset Progress

One of our most ambitious datasets yet
Completely regenerated from v1.0 — ✓ Complete Male → Female 100% v1.0 → v2.0 Regeneration 100% Scenes rewritten from scratch 1,327 🩸 The Architects Who built it — and who to blame. FB FrenzyBiscuit — Lead GD GECFDO — Generation SD Sleep Deprived — In Their Vision

🖤 License & Usage

⚠️ Usage Agreement: By using this model, you accept all terms and conditions below in full.

• 🕷️ Built upon the Apache 2.0 license
• 🔞 You confirm you are at least 18 years old
• 🩸 You accept full responsibility for all outputs generated
• 🥀 Creators bear no responsibility for how the model is used
• 🖤 Personal use only — non-profit / non-commercial

Apache 2.0 18+ Content Personal Use Only NSFW Explicit