KansenSakura-Zero-RP-12b

🌸 Techno-Organic Roleplay Engine

When the first sakura petal touched the machine, Patient Zero awoke. This narrative engine transforms stories into living infections - where every character preserves their core essence while undergoing beautiful corruption. Will your tale contain the outbreak... or become its vector?

🔍 Overview

KansenSakura-Zero-RP-12b is a roleplaying specialist model engineered for immersive narrative experiences blending Japanese visual novel aesthetics with techno-organic horror. Designed as the "Patient Zero" of narrative infection engines, it transforms characters while preserving their core essence - whether organic or mechanical.

ℹ️ Model Details

• 🧬 Core Infection: Cherry blossom motif meets nanite corruption
• ⚙️ Architecture: 12B parameter layer-merged transformer
• 🧪 Creation Method: Precision layer merging (8-model synthesis)
• 🎭 Specialization: Character-driven narratives with emergent corruption themes
• 🔖 Version: Zero (Initial Outbreak)

🎮 Intended Use

• 🤖 Character-driven narratives with transformation arcs
• 🎴 Visual novel / Doujin-style storytelling
• ☠️ Apocalyptic and cyber-horror scenarios
• 💞 Emotional corruption/redemption narratives

😷 Ethical Quarantine

This model contains:

• ⚠️ Unfiltered creative output
• ⚠️ Potential for disturbing narratives
• ⚠️ NSFW-capable layers

✍🏻 Inference Tips

1. Temperature: 0.8
2. Repetition Penalty: 1.05
3. TOP_P: 0.97
4. TOP_K: 0 (disable)
5. MIN_P: 0.025
6. Template Format: ChatML
7. Max Output: 320
8. Context Management: 16K for best quality, expect slight degradation afterwards

🧩 Model Composition

A precision surgical merge of specialized models:

Merge Details

Merge Method

This model was merged using the Passthrough merge method.

Models Merged

The following models were included in the merge:

• PocketDoc/Dans-PersonalityEngine-V1.3.0-12b
• PocketDoc/Dans-SakuraKaze-V1.0.0-12b
• elinas/Chronos-Gold-12B-1.0
• PygmalionAI/Eleusis-12B
• ReadyArt/Forgotten-Abomination-12B-v4.0
• Epiculous/Crimson_Dawn-v0.2
• LatitudeGames/Wayfarer-12B
• LatitudeGames/Muse-12B

Reproduction steps

1. Retokenize ReadyArt/Forgotten-Abomination-12B-v4.0 using mergekit-tokensurgeon

  mergekit-tokensurgeon "ReadyArt/Forgotten-Abomination-12B-v4.0" "Epiculous/Crimson_Dawn-v0.2" ./retokenized_FA --approximation-method omp --k 256

Note: After experimenting I discovered that PocketDoc/Dans-PersonalityEngine-V1.3.0-12b works with ChatML tokenizer without implicit retokenization, but produces much more text than desired. As it's position is in the starting layers, this might be a desired, more unhinged behaviour ,so we retokenize only ReadyArt/Forgotten-Abomination-12B-v4.0 to use ChatML. As we will merge it with Epiculous/Crimson_Dawn-v0.2 it's natural we use this model as a donor.

Note: according to following paper using omp --k 64 is enough and higher quantity has diminishing returns, but I decided to max the quality anyway.

2. Merge models using mergekit mergekit-multi

  name: knowledge_core
  merge_method: nuslerp
  models:
    - model: elinas/Chronos-Gold-12B-1.0
      parameters: 
        weight: 0.4
    - model: PygmalionAI/Eleusis-12B
      parameters: 
        weight: 0.6
  ---
  name: rp_blend
  merge_method: nuslerp
  models:
    - model: ./retokenized_FA
      parameters: 
        weight: 0.6
    - model: Epiculous/Crimson_Dawn-v0.2
      parameters:  
        weight: 0.4
  ---
  merge_method: passthrough
  slices:
  - sources:  # Personality Foundation
    - model: PocketDoc/Dans-PersonalityEngine-V1.3.0-12b
      layer_range: [0, 5]

  - sources:  # Base Model
    - model: PocketDoc/Dans-SakuraKaze-V1.0.0-12b
      layer_range: [5, 14]

  - sources: # Worldbuilding focus
    - model: knowledge_core
      layer_range: [14, 22]

  - sources: # Emotional intensity
    - model: rp_blend
      layer_range: [22, 29]

  - sources:  # Danger Specialization
    - model: LatitudeGames/Wayfarer-12B
      layer_range: [29, 35]

  - sources:  # Delivery & Alignment
    - model: LatitudeGames/Muse-12B
      layer_range: [35, 39]

  - sources: # Output Layer
    - model: PocketDoc/Dans-SakuraKaze-V1.0.0-12b
      layer_range: [39, 40]
  dtype: bfloat16

mergekit-multi sakuramerge.yml --intermediate-dir ./intermediates --out-path ./KansenSakura-Zero-RP-12b

Note: According to this paper top 3 layers provide up to 30% of model performance. According to this paper more complex concepts emerge in later layers. According to this paper model alignment and data presentation is most affected by last (bottom) layers. Based on this knowledge I placed different models in places where they would benefit the model the most.

3. Optional - create Q8_0 GGUF using llama.cpp

• use convert_hf_to_gguf.py script from llama.cpp (here's source)

python convert_hf_to_gguf.py ~/projects/FrankenDans-PersonalityPatchwork-VX-12b --outtype q8_0

🙏 Acknowledgments

We stand on the shoulders of giants:

• PocketDoc for PersonalityEngine and SakuraKaze foundations
• Latitude team for narrative expertise
• Elinas for temporal knowledge systems
• PygmalionAI for emotional intelligence research
• ReadyArt for dark arts
• Arcee AI for making questionable AI combinations possible with mergekit
• Featherless AI for kindly hosting the model
• Team mradermacher for awesome quants
• You, dear user, for willingly exposing yourself to this digital infection vector. Patient Zero status granted! 🦠

When the first circuit blooms... the infection begins

📜 Narrative Hazard Disclaimer

KansenSakura-Zero-RP-12b is provided "as found in the corrupted data-core" without warranty of any kind. Users assume all responsibility for unintended character corruptions, emergent techno-organic fantasies, or sudden urges to describe rusting cherry blossoms. Not approved for medical diagnostics, financial advice, or anti-zombie defense systems. May contain traces of actual emotional intelligence. Side effects may include: phantom nanite tingling, involuntary haiku composition, or temporary possession by tragic android protagonists. If worldbuilding symptoms persist for more than 4 narrative hours, consult your nearest cyber-shaman. Remember: This isn't an infection - it's a feature.

Disclaimer v1.0 - Valid until next bloom cycle 🌸⚙️💀

New model available