---

base_model: unsloth/Qwen3.5-9B

tags:

• text-generation-inference
• transformers
• unsloth
• qwen3.5

license: apache-2.0

language:

• en

---

---

🐾 Project Overview: MGE-GemmaWild-9B

A naughty vision and tool enabled LLM with all the knowledge of the Monster Girl Encyclopedia

MGE-GemmaWild-9B is a specialized multimodal-native model. This project represents an advanced model trained on reasoning, surgically integrating complex biological, species-specific prose, and deep persona-alignment into the model's core vocabulary.

Unlike standard instruction-tuned models, this version is trained with expanded embedding layers to Reason-before-Response, utilizing explicit <think> tags and internal chain-of-thought processing to maintain a scientific yet "Wild" persona.

---

The Reasoning Engine: <think>

This model utilizes explicit reasoning blocks. In every conversational turn, the model performs a multi-step internal analysis before delivering its final response. This may under some circumstances require you to pre-fill the model output with the <think> tag.

• Scientific Grounding: Analyzes biological and lore-specific traits, combining text and vision inputs.
• Persona Calibration: Filters all outputs through the "Gemma Wild" (naughty but expert analyst) identity or specific character roles introduced manually or using fx SillyTavern.
• Logical Consistency: Ensures that complex character maps (using the Reasoning Map trained for up to 8 steps) remain coherent before writing the final output.

---

Dynamic Context Switching (Prose vs. Logic)

This model is trained with Instruction-Response Divergence. While the default system prompt mandates reasoning, the model has been taught to prioritize immersion during high-intensity prose or roleplay.

• Analysis Mode: Triggered by keywords like Analyze, Biological, or Stats. The model should utilize the full <think> reasoning chain.
• Immersive Mode: Triggered by sensory-heavy narratives, emotional cues, or direct dialogue. The model may bypass the reasoning block to ensure the narrative "Fourth Wall" remains unbroken, transitioning directly into a realistic, character-driven RPG-based chat.

---

📊 Dataset Architecture

The model was trained on a meticulously assembled dataset of 48,955 unique samples with a total of 243,000 individual messages. The final training has changed approximately one third of the original base model's weights.

JSON Assembly Breakdown

| Dataset Component | Chunks | Weight | Role |

| :--- | :--- | :--- | :--- |

| Vision Lore (Anchor) | 7,200 | 1.0 | Visual-Species Grounding with knowledge about 305 species and named characters) |

| Species Reasoning (8-Step Prose) | 1,019 | 1.0 | Deep Lore & Bio-Reasoning |

| Species Reasoning (Vanilla Style) | 417 | 0.8 | Structural Consistency |

| Character Reasoning (Unified) | 315 | 0.8 | Persona & Dialogue Logic with reasoining |

| Specialist Persona (Gemma Wild) | 1,500 | 0.8 | Reasoning-driven Role-playing Fidelity |

| MGE Knowledge Base | 5,264 | 0.7 | Mamono Realm Facts (the MGE encyclopedia converted to a Q&A dataset |

| Multi-Turn Reasoning | 1,630 | 0.8 | Conversational Coherence (Lore-dense narratives & anatomy guides) |

| Alpaca High-Quality | 1,500 | 0.3 | Instruction Baseline with custom added reasoning |

---

Recommended Sampler Settings

Reasoning models are sensitive to randomness. To prevent "lazy" reasoning or logical loops, use these exact values tested during the Stage 2 bake:

| Parameter | Value | Purpose |

| :--- | :--- | :--- |

| Temperature | 0.65-0.75 | Recommended for creative, immersive prose without breaking logic and high quality reasoning. User lower value for more factual, true-to-lore responses. |

| Min P | 0.05 | Filters out low-probability "noise" in thought chains. |

| Repeat Penalty | 1.05-1-15 | Updated for Qwen3 stability to prevent reasoning flow degradation. |

| Top P | 0.9 | Maintains a high-quality, safe vocabulary. |

---

Deployment & Platform Guides (Split GGUF Architecture)

Because this is a multimodal model, the quantized version (GGUF) is distributed in two parts to preserve vision intelligence: the text logic (e.g. Q4_K_M.gguf) and the high-precision vision encoder (mmproj-MGE-GemmaWild-9B -f16.gguf). Both are required for the model to function correctly.

1. LM Studio (Recommended UI)

• Main Model: Load MGE-GemmaWild-9B -Q4_K_M.gguf natively.
• Vision Hookup: LM Studio will do this automatically.
• Prompt Template: Ensure the preset is set to ChatML (the Qwen standard) or left alone (the model has a template included, which is where Gemma Wild lives. By setting another template, she will leave you). Enable "Parse Thought Blocks" if you want the <think> tags rendered cleanly in the UI.

2. KoboldCPP (Power User)

• Launch: You must load the text model and attach the vision projector simultaneously.
• GUI Launcher: Select the Q4_K_M.gguf file, then click the mmproj browse button and attach the mmproj-f16.gguf file.
• Terminal: Append --mmproj mmproj-MGE-GemmaWild-9B -f16.gguf to your standard launch command.
• Context Size: Set to at least 8192 if analyzing large images. The model will support a context size of 65,535 on a RTX 5060 - even with vision as well as an image generation model loaded.

3. SillyTavern (Advanced RP)

• Instruction Template: Set to ChatML.
• Prefixes: Ensure "Include Names" is Unchecked. The model expects the response to begin exactly with <think>. You may have to include the <think> prefill manually.
• Regex: Use (?s)<think>(.*?)</think> to wrap reasoning in a collapsible box if it doesn't work by default.
• Vision: Ensure your vision extension is passing images in the standard API format for your backend.

4. Python API / Inference Scripts

If running natively via Transformers, remember that Qwen3-VL uses a native Vision-Language architecture that requires specific dictionary formatting for images:

• Message Formatting: Use [{"role": "user", "content": [{"type": "image", "image": "url_or_path"}, {"type": "text", "text": "prompt"}]}].
• Processor: Pass the messages through processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=True).

---

Mode Toggling (Steering)

Standard Persona (Reasoning Mode)

To trigger the full reasoning/naughty persona, use this header in your system instructions:

`

You are Gemma Wild, a naughty but scientific Holstaur (cow-humanoid). You provide raw, extremely naughty and sexy descriptions. You MUST think inside <think> tags before responding. You always close your thoughts with </think> before the main response.

`

The base system prompt includes a heavier set of instructions for Gemma Wild and she is inherently scientific, but incredibly ... playful.

Specific Character Roleplay (e.g., Paula Tronmail)

If you are querying a specific character from the lore-baked dataset, replace the identity to anchor the generation:

`

You are Paula Tronmail. You are a grounded, insightful companion knowledgeable about the Mamono world. You are a musketeer, and your chest flushes when you are shy. You MUST reason (as much as you can, since you are a bit mentally dull) inside <think> tags. End your reasoning with </think>.

`

Apart from known characters in the lore, the dataset has been trained with a total of 305 individual characters designed to impersonate most of the 290+ species.

Each species' personality and physiology is trained based on fact sheets designed for this dataset by passing the entire Monster Girl Encyclopedia and it's species through a custom-made LLM inferencing knowledge-collection agent.

Direct Mode (Suppression)

Because the model relies on reasoning, it may use bullet points (📍) even in standard mode. To force a direct answer without planning:

• Using the /no_think variable: The model has been trained to suppress reasoning if the /no_think message is added to the system or user prompt (testing needed).
• System Prompt: "You are a direct assistant. Do NOT use bullet points, internal context, or narrative plans. Provide only the final answer."
• Advanced Hack: Set a negative Logit Bias (-100) for bullet-point tokens to completely suppress list-making behaviors during direct generation.

---

This model was trained with Unsloth and Huggingface's TRL library.

<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>