Darkhn-Quants-4/Gemma-4-31B-Animus-V15.0-GGUF overview

---

license: apache-2.0

tags:

• finetune
• roleplay
• chat
• wings-of-fire
• nsfw
• not-for-all-audiences

base_model:

• Darkhn/Gemma-4-31B-Animus-V15.0

base_model_relation: quantized

---

<style>

body {

font-family: 'Quicksand', sans-serif;

background: linear-gradient(135deg, #4a1e00 0%, #1c0a00 100%);

color: #F5EFE6;

margin: 0;

padding: 0;

font-size: 16px;

}

h1, h2, h3, h4, summary {

font-family: 'Cinzel', serif;

}

.container {

margin: 20px auto;

max-width: 900px;

background-color: rgba(28, 22, 18, 0.95);

padding: 30px;

border-radius: 12px;

box-shadow: 0 4px 20px rgba(255, 140, 0, 0.15);

border: 1px solid rgba(255, 140, 0, 0.2);

outline: 1px solid rgba(255, 140, 0, 0.5);

outline-offset: -1px;

position: relative;

}

.container::before {

content: '';

position: absolute;

top: -1px;

left: -1px;

right: -1px;

bottom: -1px;

border: 1px solid rgba(255, 165, 0, 0.98);

border-radius: 12px;

pointer-events: none;

animation: borderGlow 2.5s ease-in-out infinite;

}

@keyframes borderGlow {

0% {

box-shadow: 0 0 5px rgba(255, 165, 0, 0.98);

}

50% {

box-shadow: 0 0 12px rgba(255, 165, 0, 0.98);

}

100% {

box-shadow: 0 0 5px rgba(255, 165, 0, 0.98);

}

}

.header h1 {

font-size: 32px;

color: #FFA500;

margin: 0 0 20px 0;

text-align: center;

text-shadow: 0 0 12px rgba(255, 100, 0, 0.6);

}

.info img {

width: 100%;

max-width: 700px;

display: block;

margin: 0 auto 25px auto;

border-radius: 10px;

box-shadow: 0 0 20px rgba(255, 140, 0, 0.25);

border: 1px solid rgba(255, 140, 0, 0.2);

outline: 1px solid rgba(255, 140, 0, 0.5);

outline-offset: -1px;

}

a {

color: #FFD700;

text-decoration: none;

transition: color 0.3s ease;

}

a:hover {

color: #FFDAB9;

}

.button {

display: inline-block;

background-color: #E55B00;

color: #FFFFFF;

padding: 12px 24px;

border-radius: 5px;

cursor: pointer;

text-decoration: none;

font-family: 'Cinzel', serif;

font-weight: 600;

transition: all 0.3s ease;

border: 1px solid transparent;

margin-right: 10px;

margin-bottom: 10px;

}

.button:hover {

background-color: #FF8C00;

box-shadow: 0 0 15px rgba(255, 140, 0, 0.5);

transform: translateY(-2px);

}

pre {

background-color: rgba(45, 35, 25, 0.95);

padding: 15px;

border-radius: 5px;

overflow-x: auto;

border: 1px solid rgba(255, 140, 0, 0.2);

outline: 1px solid rgba(255, 140, 0, 0.5);

outline-offset: -1px;

}

code {

font-family: 'Courier New', monospace;

color: #F5EFE6;

}

.section-container {

margin: 40px 0;

}

h2 {

font-size: 26px;

color: #FFA500;

text-shadow: 0 0 10px rgba(255, 140, 0, 0.5);

border-bottom: 1px solid rgba(255, 140, 0, 0.2);

padding-bottom: 10px;

margin-bottom: 20px;

}

.info-card {

background: rgba(45, 35, 25, 0.95);

border: 1px solid rgba(255, 140, 0, 0.2);

border-radius: 8px;

overflow: hidden;

margin-bottom: 25px;

}

.info-header {

background: rgba(255, 140, 0, 0.1);

padding: 20px;

border-bottom: 1px solid rgba(255, 140, 0, 0.2);

}

.info-header h3 {

color: #FFA500;

margin: 0 0 10px 0;

font-size: 22px;

text-shadow: 0 0 5px rgba(255, 140, 0, 0.3);

}

.model-tags {

display: flex;

gap: 8px;

flex-wrap: wrap;

}

.model-tag {

background: rgba(218, 165, 32, 0.15);

color: #FFD700;

padding: 4px 8px;

border-radius: 4px;

font-size: 12px;

border: 1px solid rgba(218, 165, 32, 0.3);

font-family: 'Quicksand', sans-serif;

}

.card-content {

padding: 20px;

line-height: 1.7;

}

.card-content p, .card-content li {

margin-bottom: 1em;

}

.card-content p:last-child, .card-content li:last-child {

margin-bottom: 0;

}

.card-content ul {

list-style: none;

padding-left: 20px;

}

.card-content li::before {

content: '✦';

color: #FFD700;

font-weight: bold;

display: inline-block;

width: 1em;

margin-left: -1.2em;

font-size: 1.2em;

line-height: 1;

}

.card-content strong {

color: #FFD700;

font-weight: 600;

}

.config-container {

background: rgba(45, 35, 25, 0.95);

border: 1px solid rgba(255, 140, 0, 0.2);

border-radius: 8px;

overflow: hidden;

}

.config-header {

background: rgba(255, 140, 0, 0.1);

padding: 15px 20px;

border-bottom: 1px solid rgba(255, 140, 0, 0.2);

}

.config-header h3 {

margin: 0;

color: #FFA500;

font-size: 22px;

}

.config-content {

padding: 20px;

display: grid;

grid-template-columns: repeat(auto-fit, minmax(200px, 1fr));

gap: 20px;

}

.config-item {

display: flex;

flex-direction: column;

gap: 5px;

}

.config-label {

color: #FFD700;

font-size: 14px;

font-weight: 500;

font-family: 'Quicksand', sans-serif;

}

.config-value {

color: #F5EFE6;

font-family: 'Courier New', monospace;

font-size: 18px;

font-weight: bold;

}

.link-arrow {

display: inline-block;

transition: transform 0.3s ease;

}

a:hover .link-arrow {

transform: translateX(3px);

}

.support-section {

text-align: center;

margin-top: 40px;

background: rgba(45, 35, 25, 0.95);

border: 1px solid rgba(255, 140, 0, 0.2);

border-radius: 8px;

padding: 20px;

}

.support-section p {

margin-bottom: 15px;

font-size: 1.1em;

margin-top: 0;

}

summary {

cursor: pointer;

list-style: none;

outline: none;

display: flex;

align-items: flex-start;

}

summary::-webkit-details-marker {

display: none;

}

summary::before {

content: '▶';

font-size: 1.2em;

color: #FFA500;

margin-right: 15px;

padding-top: 5px;

transition: transform 0.2s ease;

flex-shrink: 0;

}

details[open] > summary::before {

transform: rotate(90deg);

}

summary > h2 {

flex-grow: 1;

}

.format-list {

margin: 1.5em 0;

}

.format-list dt {

color: #FFA500;

font-family: 'Cinzel', serif;

font-weight: 600;

font-size: 1.3em;

text-shadow: 0 0 8px rgba(255, 140, 0, 0.5);

margin-bottom: 0.6em;

}

.format-list dd {

margin-left: 0;

margin-bottom: 1.5em;

background-color: rgba(28, 22, 18, 0.95);

padding: 12px 15px;

border-radius: 6px;

border: 1px solid rgba(255, 140, 0, 0.2);

border-left: 4px solid #FFA500;

}

.format-list dd:last-of-type {

margin-bottom: 0;

}

</style>

<div class="container">

<link href="https://fonts.googleapis.com/css2?family=Cinzel:wght@400;500;600&family=Quicksand:wght@400;500&display=swap" rel="stylesheet">

<div class="header">

<h1>Gemma-4-31B-Animus-V15.0-GGUF</h1>

</div>

<div class="info">

<img src="image.png" alt="Wings_of_Fire" width="700">

<div class="support-section">

<p><strong>Send me your support to help me feed the data beast! also taking comissions for universe specific models</strong></p>

<a href="https://ko-fi.com/som1tokmynam" target="_blank" class="button">

Support on Ko-fi

</a>

</div>

<div class="section-container">

<details open>

<summary><h2>Critical: Chat Template & Setup</h2></summary>

<div class="info-card">

<div class="card-content">

<p>This model uses the standard <strong>Gemma 4</strong> chat template. The structure is complex and requires specific configuration for correct performance. Please read the following sections carefully.</p>

<p><strong>Human-Readable Format:</strong></p>

<pre><code>&lt;bos&gt;&lt;|turn&gt;system

&lt;|think|&gt;

{system prompt}&lt;turn|&gt;

&lt;|turn&gt;user

{user message}&lt;turn|&gt;

&lt;|turn&gt;model

&lt;|channel&gt;thought

{reasoning}&lt;channel|&gt;{response}&lt;turn|&gt;</code></pre>

</div>

</div>

<div class="info-card">

<div class="info-header"><h3>⚠️ Important: Use Chat Completion, Not Text Completion</h3></div>

<div class="card-content">

<p><strong>Strongly recommended:</strong> Always use the <code>/v1/chat/completions</code> endpoint. Correct formatting is handled automatically by the server.</p>

<p>Text completion (<code>/v1/completions</code>) is currently broken for impersonations without a custom patch. Sending the context as a user prompt and asking the model to generate the response can produce a malformed turn structure.</p>

<p><strong>Reasoning with text completion and Llama.cpp specific: </strong>You must prefill the response with <code>start_reply_with: <|channel>thought + \n</code> (a new line) or reasoning will not trigger (you will get blank responses if you enabled reasoning and did not prefill)</p>

</div>

</div>

<div class="info-card">

<div class="info-header"><h3>Chat Completion API Configuration (SillyTavern, etc.)</h3></div>

<div class="card-content">

<p>When using an OpenAI-compatible API, you must add custom parameters to control thinking and ensure correct token handling. In frontends like <strong>SillyTavern</strong>, add the following JSON to the <strong>"Custom Parameters"</strong> (or <code>extra_body</code>) field in your connection settings:</p>

<pre><code>chat_template_kwargs: {enable_thinking: true,

preserve_thinking: false}

skip_special_tokens: false

</code></pre>

<ul>

<li><code>enable_thinking: true</code>: Activates the model's chain-of-thought reasoning. Change to <code>false</code> to disable it for faster, more direct responses.</li>

<li><code>skip_special_tokens: false</code>: <strong>This is critical.</strong> It prevents the API from stripping away the special tokens (like <code>&lt;|channel&gt;</code> and <code>&lt;channel|&gt;</code>) that are required for the model's chat template to work correctly.</li>

<li><code>preserve_thinking: false</code>: <strong>Optional, Default: false.</strong> newly added to the chat template, It allows reusing previous reasoning turns in </li>

</ul>

</div>

</div>

<div class="info-card">

<div class="info-header"><h3>General Notes</h3></div>

<div class="card-content">

<ul>

<li>The model retains the vision adapter from the base Gemma 4 31B and supports image inputs.</li>

<li>Both thinking and no-thinking modes were included in training data — use <code>enable_thinking</code> to select at runtime.</li>

<li>The <code><|think|></code> token in the system turn is what activates chain-of-thought reasoning at the architecture level.</li>

</ul>

</div>

</div>

</details>

</div>

<div class="section-container">

<details>

<summary><h2>Quantized Models</h2></summary>

<div class="info-card">

<div class="card-content">

<p>The quantized model files are available for download. Click the button below to view the files.</p>

<a href="https://huggingface.co/Darkhn-Quants-4/Gemma-4-31B-Animus-V15.0-EXL3" target="_blank" class="button">

Download EXL3 Files <span class="link-arrow">→</span>

</a>

</div>

</div>

</details>

</div>

<div class="section-container">

<details>

<summary><h2>Character Card & Lore Book</h2></summary>

<div class="info-card">

<div class="card-content">

<p>For the best roleplaying experience, it is highly recommended to use the provided character card and lore book. These files help guide the model's persona and provide rich, in-universe context.</p>

<a href="https://huggingface.co/Darkhn/Sampler_settings_and_system_prompt/tree/main/character_card" target="_blank" class="button">

Download Files <span class="link-arrow">→</span>

</a>

</div>

</div>

</details>

</div>

<div class="section-container">

<details>

<summary><h2>Sampler Presets</h2></summary>

<div class="info-card">

<div class="card-content">

<p>For a seamless setup in SillyTavern, you can download pre-configured sampler presets. These are tuned to provide an optimal balance between creativity and narrative coherence for this model.</p>

<p>Simply download the <code>.json</code> file below and import it into SillyTavern's sampler presets menu.</p>

<a href="https://huggingface.co/Darkhn/Sampler_settings_and_system_prompt/raw/main/Gemma_4_Think_SillyTavern_settings.json" target="_blank" class="button">

Download SillyTavern Preset (Thinking) <span class="link-arrow">→</span>

</a>

<a href="https://huggingface.co/Darkhn/Sampler_settings_and_system_prompt/raw/main/Gemma_4_NoThink_SillyTavern_settings.json" target="_blank" class="button">

Download SillyTavern Preset (No Thinking) <span class="link-arrow">→</span>

</a>

<p><li>For those that dont use silly tavern, Samplers settings are:</li></p>

<ul>

<p><strong>Temperature:</strong> 1.0</p>

<p><strong>Min P:</strong> 0.02</p>

</ul>

</div>

</div>

</details>

</div>

<div class="section-container">

<details open>

<summary>

<h2>Roleplay Format Guide</h2>

</summary>

<div class="info-card">

<div class="card-content">

<p>For the best results, use this structured format. This helps the AI clearly distinguish between actions, inner thoughts, dialogue, and — when reasoning is enabled — narrative planning.</p>

<dl class="format-list">

<dt>Actions / Descriptions</dt>

<dd><code>He walked across the room and stared out the window.</code></dd>

<dt>Inner Thoughts</dt>

<dd><code>-I wonder what she's thinking.-</code></dd>

<dt>Dialogue</dt>

<dd><code>Alex (Curious): "What do you see out there?"</code></dd>

<dt>Reasoning / Narrative Planning (CoT)</dt>

<dd>

<p>When the reasoning channel is active, the model uses it as a <strong>director's notebook</strong> — planning what should happen and why, <em>before</em> narrating the scene. This is CYOA/GM-style intent planning, not deliberative self-critique.</p>

<p><strong>Example:</strong></p>

<pre><code>*The user chose to open the door despite my warnings.

I'll reward this bold choice by revealing something

unexpected — not a monster, but an empty room, letting

the tension build through absence rather than action.

Sensory focus: dust, stale air, distant echo.*

</code></pre></dd>

<h4>Reasoning traces are intentionally short (~100–200 tokens)</h4>

</div>

</div>

</details>

</div>

<div class="section-container">

<details>

<summary><h2>Roleplay Example</h2></summary>

<div class="info-card">

<div class="card-content">

<p>Click the button below to view a full, unedited chatlog demonstrating the model's narrative style and character portrayal.</p>

<a href="https://huggingface.co/Darkhn/Sampler_settings_and_system_prompt/blob/main/chat_logs/chatlog_mistral_large.png" target="_blank" class="button">

View Chatlog Example <span class="link-arrow">→</span>

</a>

</div>

</div>

</details>

</div>

<div class="section-container">

<details open>

<summary><h2>Model Description</h2></summary>

<div class="info-card">

<div class="card-content">

<p>This is <strong>Version 15.0</strong> in the Animus series. Built on <strong>google/gemma-4-31B-it</strong>, this release features a significantly expanded dataset and revised training composition compared to prior versions.</p>

<p>The Supervised Fine-Tuning (SFT) dataset consists of <strong>122 million tokens</strong> across <strong>22,000 samples</strong>, structured to balance standard interactions, formatting adherence, and complex narrative paths:</p>

<ul>

<li><strong>Standard Roleplay &amp; Lore (12,000 samples):</strong> Focused on in-character interactions, dialogue, and general world-building.</li>

<li><strong>NSFW Scenarios (4,000 samples):</strong> Included to improve narrative flexibility in mature or adult-themed contexts.</li>

<li><strong>Bad Endings (4,000 samples):</strong> Designed to handle darker themes, conflict resolution, and complex narrative outcomes.</li>

<li><strong>Instruction Formatting (2,000 samples):</strong> Q&amp;A pairs tailored to improve structured task adherence.</li>

</ul>

<p>Additionally, Version 15.0 incorporates <strong>71 million tokens of Chain of Thought (COT) reasoning</strong> across <strong>16,000 samples</strong>. These reasoning blocks function similarly to "dungeonmaster notes," helping the model plan logical progression and thematic consistency behind the scenes before outputting responses.</p>

<p>The combined dataset aims to provide steady prose quality, better thematic consistency, and an improved grasp of in-universe lore.</p>

<p><em>Formatting Note:</em> The model remains highly sensitive to the formatting of the system prompt and the initial turn. If the first assistant message is brief, subsequent responses will generally match that length.</p>

</div>

</div>

</details>

</div>

<div class="info-card">

<div class="info-header">

<h3>V15.0 Training Process</h3>

</div>

<div class="card-content">

<p>

V15.0 was fine-tuned using a <strong>two-stage LoRA recipe on disjoint datasets</strong>,

designed to first establish a stable prose/narration baseline, then add a

narrative-planning reasoning channel on top without disturbing the Phase 1 behavior.

All training was performed locally on consumer hardware.

</p>

<p>

After phase 1 was completed, modelstock was executed on the checkpoints, using epoch 0.4 as the base, and every subsequent checkpoints as the stock, that merged model was then used as the base for phase 2.

</p>

<ul>

<li><strong>Base Model:</strong> <code>google/gemma-4-31B-it</code></li>

<li><strong>Hardware:</strong> 4 × RTX 3090 (local workstation)</li>

<li><strong>Framework:</strong> Axolotl</li>

<li><strong>Total training time:</strong> ~362 hours (~15 days) of wall-clock compute</li>

</ul>

<h3>Phase 1 — SFT (Prose &amp; Instruction-Following)</h3>

<ul>

<li><strong>LoRA rank:</strong> 128</li>

<li><strong>Epochs:</strong> 2</li>

<li><strong>Training time:</strong> ~290 hours</li>

</ul>

<h3>Phase 2 — CoT Specialization (Narrative Reasoning)</h3>

<ul>

<li><strong>LoRA rank:</strong> 32</li>

<li><strong>Epochs:</strong> 1</li>

<li><strong>Training time:</strong> ~72 hours</li>

</ul>

</div>

</div>

<div class="section-container">

<details>

<summary><h2>Intended Use & Limitations</h2></summary>

<div class="info-card">

<div class="card-content">

<ul>

<li><strong>Intended Use:</strong> The primary purpose of this model is for creative roleplay within the <em>Wings of Fire</em> universe. However, user feedback indicates it is also highly effective for general-purpose roleplaying.</li>

<li><strong>Limitations & Quirks:</strong>

<ul>

<li>Performance on tasks outside of its training domain (general knowledge, coding, etc.) is not guaranteed and will likely be poor.</li>

<li><strong>Versatility:</strong> While it appears to be only a <em>Wings of Fire</em> tuned model, users have reported it is very capable of performing normal roleplay with other settings and characters.</li>

<li><strong>Dragon Anatomy Bias (Expected Behavior):</strong> The model has a strong prior toward <em>draconic</em> anatomy — wings, claws, talons, scales, snouts, tails — even when roleplaying non-dragon characters. This is <strong>intentional and expected</strong>: human-anatomy terms (hands, fingers, feet, hair, skin, etc.) were heavily filtered out of the training data to preserve the <em>Wings of Fire</em> setting's integrity. If you want the model to write human or humanoid characters cleanly, provide explicit anatomical cues in the character card or system prompt (e.g. <em>"a human with hands and fingers, no wings or claws"</em>), otherwise expect draconic descriptors to leak through.</li>

<li>The model may "hallucinate" or generate plausible but non-canonical information, especially when pushed outside the established "what-if" scenarios.</li>

<li><strong>Content:</strong> The training data includes mature and darker themes from the <em>Wings of Fire</em> series, such as conflict, character death, and moral ambiguity. The model is capable of generating content reflecting these themes. As always, it is up to the user what they do with it.</li>

<li><strong>Formatting:</strong> Training data was cleaned to remove narrative artifacts like <code>scene transitions</code>. The model should now produce cleaner prose.</li>

<li><strong>Safety:</strong> This model has not undergone additional safety alignment beyond what was included in its base model. Standard responsible AI practices should be followed.</li>

</ul>

</li>

</ul>

</div>

</div>

</details>

</div>

<div class="section-container">

<details>

<summary><h2>Acknowledgements</h2></summary>

<div class="info-card">

<div class="card-content">

<ul>

<li>Credit to Google for the powerful gemma-4-31B-it model.</li>

<li>Credit to Google for the Gemini Pro model, used in dataset generation.</li>

<li>Credit to Anthropic for sonnet 4.5, used in dataset generation.</li>

<li>Credit to Hangzhou DeepSeek Artificial Intelligence for the deepseek model, used in dataset generation.</li>

<li>Credit to Moonshot AI for the Kimi K2 model, used in dataset generation.</li>

</ul>

</div>

</div>

</details>

</div>

</div>

</div>