KansenSakura-Zero-RP-12b

Rusted petals fall
On circuits that dream of blood
Error 0x1FABE5: Beauty not found

This is not a bug
It's the feature they warned of
Reboot into spring

🌸 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

Temperature: 0.8
Repetition Penalty: 1.05
TOP_P: 0.97
TOP_K: 0 (disable)
MIN_P: 0.025
Template Format: ChatML
Max Output: 320
Context Management: 16K for best quality, expect slight degradation afterwards

🧩 Model Composition

A precision surgical merge of specialized models:

Layer Range	Model	Contribution
0-5	`Dans-PersonalityEngine-V1.3.0`	Personality anchoring
5-14	`Dans-SakuraKaze-V1.0.0`	Narrative coherence
14-22	`Chronos-Gold-12B` + `Eleusis-12B`	World knowledge & emotional intelligence
22-29	`Forgotten-Abomination-12B-v4.0` + `Crimson_Dawn-V0.2`	RP memory & corruption mechanics
29-35	`Wayfarer-12B`	Scene crafting
35-39	`Muse-12B`	Immersive delivery
39-40	`Dans-SakuraKaze-V1.0.0`	Output coherence

Merge Details

Merge Method

This model was merged using the Passthrough merge method.

Models Merged

The following models were included in the merge:

Reproduction steps

Spoiler warning

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.

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.

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
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 🌸⚙️💀

Retreatcost
/

KansenSakura-Zero-RP-12b