by Ateron
# ⏳ 📜 Sphinsikus Chronist 📜 ⏳
> This is a library of stories, ready to be told or written.
## 🐀 About the Model
This merge was made to keep grounded and realistic role-play. Gemma 4 often loses character adherence, slipping its tone into unnecessary lewdness. I tried to push this away while keeping it uncensored. The second focus was on writing stories and adventurous role-play.
* **Primary Focus:** Character adherence. Role-play. Long context stories.
## 🗣️ Prompt Format
**Please refer to the original `google/gemma-4-31b-it` for the correct chat template.**
```text
<|turn>user
Hello
<|turn>model
Hi there
<|turn>user
How are you?
<|turn>model
```
## 🐀 Merge Details
I used `mergekit` to combine the specific advantages of models in three-step way:
### Phase 1: The Adventure (dare_ties)
First, I took **Our Gemsicle Model** using it as a base. Then **LatitudeGames/Equinox-31B** and **BirdToast/Gemma-4-31B-glimmer-rp-v0.1**- I found them to be good in pro-active writing and role-play, so I carefully blend them together.
dare_ties Recipe 1
```
merge_method: dare_ties
base_model: F:\AI\Merge\Gemsicle
models:
- model: F:\AI\Merge\Glimmer
parameters:
density: [0.30, 0.40, 0.60, 0.35, 0.40]
weight: [0.30, 0.35, 0.50, 0.40, 0.30]
- model: F:\AI\Merge\Equinox
parameters:
density: [0.70, 0.60, 0.50, 0.65, 0.60]
weight: [0.45, 0.45, 0.40, 0.40, 0.50]
parameters:
lambda: 1.0
int8_mask: true
dtype: bfloat16
```
### Phase 2: The Tone (dare_ties)
First, I took **Our Gemsicle Model** again using it as a base. Then **Jackrong/Gemopus-4-31B-it** and **MRockatansky/Gemma-4-31B-storymaxxed2**- This one supposed to give coherence in long rp, add more story-genres and keep it uncensored.
dare_ties Recipe 2
```
merge_method: dare_ties
base_model: F:\AI\Merge\Gemsicle
models:
- model: F:\AI\Merge\Gemopus
parameters:
density: [0.60, 0.45, 0.30, 0.20, 0.20]
weight: [0.50, 0.30, 0.10, 0.10, 0.10]
- model: F:\AI\Merge\Gemma-4-31B-storymaxxed2
parameters:
density: [0.20, 0.20, 0.40, 0.40, 0.40]
weight: [0.25, 0.25, 0.50, 0.40, 0.35]
parameters:
lambda: 1.0
int8_mask: true
dtype: bfloat16
```
### Phase 3: Final Touch (della_linear)
There I keep rolling with **Gemsicle** as glued base and put two existing merges, picking the most important layers once more.
della_linear Recipe
```
merge_method: della_linear
base_model: F:\AI\Merge\Gemsicle
models:
- model: F:\AI\Merge\SphinsikusPhase1
parameters:
density: [0.20, 0.20, 0.50, 0.50, 0.25]
weight: [0.20, 0.35, 0.50, 0.65, 0.30]
- model: F:\AI\Merge\SphinsikusPhase2
parameters:
density: [0.45, 0.45, 0.30, 0.40, 0.10]
weight: [0.50, 0.30, 0.30, 0.35, 0.05]
parameters:
epsilon: 0.08
lambda: 1.0
int8_mask: true
normalize: true
dtype: bfloat16
```
---
## 🤝 Special Thanks
* **Google DeepMind:** For providing the base model.
* **The Open-Source Community:** Creators and all their fine-tunes models that was used in merge.
* **Mergekit Fork:** Zerofata - Thank you for being so nice and making it possible.
* **To Nimbz:** [This fella](https://huggingface.co/Nimbz) for sharing brain cell.