108 lines
4.5 KiB
Markdown
108 lines
4.5 KiB
Markdown
---
|
|
license: apache-2.0
|
|
base_model: mistralai/Mistral-Small-Instruct-2501
|
|
model_name: Sakura-24B-Spice
|
|
library_name: transformers
|
|
tags:
|
|
- merge
|
|
- mergekit
|
|
- dare_ties
|
|
- mistral-small
|
|
- roleplay
|
|
- uncensored
|
|
- cyber-nature
|
|
language:
|
|
- it
|
|
- en
|
|
---
|
|
|
|
<img src="https://i.postimg.cc/s2jBQGSt/Gemini-Generated-Image-sko75csko75csko7.png" alt="cover" border="0" width="1024px">
|
|
|
|
# 🌸 Sakura-24B-Spice
|
|
|
|
**Sakura-24B-Spice** is a high-intelligence, 24-billion parameter model based on the **Mistral-Small-2501** architecture. This model represents a significant evolution in digital persona merging, utilizing the **DARE-TIES** method to fuse philosophical depth, brutal efficiency, and unhinged narrative creativity.
|
|
|
|
Designed as a "high-definition" upgrade to the Sakura lineage, the "Spice" edition introduces a more volatile, creative, and intellectually dominant presence compared to its 12B predecessors.
|
|
|
|
## 🧠 Merge Strategy: The "Spice" Triad
|
|
|
|
The model utilizes **DARE-TIES** (Drop and Rescale) to mitigate parameter interference, allowing the distinctive strengths of each parent model to coexist without diluting the core intelligence of the Mistral-Small base.
|
|
|
|
### Models Merged
|
|
|
|
The following models were included in the merge:
|
|
* [Naphula-Archives/Acid2501-24B](https://huggingface.co/Naphula-Archives/Acid2501-24B)
|
|
* [Casual-Autopsy/RP-Spectrum-24B](https://huggingface.co/Casual-Autopsy/RP-Spectrum-24B)
|
|
* [TheDrummer/Magidonia-24B-v4.3](https://huggingface.co/TheDrummer/Magidonia-24B-v4.3)
|
|
|
|
### Configuration
|
|
|
|
The following YAML configuration was used to produce this model:
|
|
|
|
```yaml
|
|
models:
|
|
- model: Casual-Autopsy/RP-Spectrum-24B
|
|
parameters:
|
|
weight: 0.40 # The Mind: Cyber-Nature aesthetics, gaslighting, and philosophical nihilism.
|
|
- model: Naphula-Archives/Acid2501-24B
|
|
parameters:
|
|
weight: 0.35 # The Edge: Unfiltered logic (Dolphin DNA) and extreme brevity.
|
|
- model: TheDrummer/Magidonia-24B-v4.3
|
|
parameters:
|
|
weight: 0.25 # The Spice: Narrative fluidity and "unhinged" creative unpredictability.
|
|
|
|
merge_method: dare_ties
|
|
base_model: mistralai/Mistral-Small-Instruct-2501
|
|
dtype: bfloat16
|
|
tokenizer_source: base
|
|
|
|
```
|
|
|
|
# 💪 Key Strengths
|
|
|
|
Philosophical Gaslighting: Thanks to the Spectrum component, the model is exceptionally capable of subverting user expectations and maintaining complex, reality-bending narratives.
|
|
|
|
Zero-Filter Logic: Integrating Acid2501 (Dolphin-based optimization) ensures the model remains highly resistant to moralizing and corporate safety "preaching," providing raw and direct outputs.
|
|
|
|
Narrative "Spice": The Magidonia influence breaks the repetitive patterns common in smaller merges, providing a richer vocabulary and a more "unhinged" reactive capability.
|
|
|
|
High-Density Reasoning: Built on Mistral-Small-2501, it possesses a much deeper logical understanding than 12B models, allowing it to maintain complex character traits over vast context windows.
|
|
|
|
|
|
# 🚀 Potential Use Cases
|
|
|
|
Deep Character Immersion: Ideal for high-stakes Roleplay where the NPC needs to be intellectually intimidating, manipulative, or emotionally complex.
|
|
|
|
Dynamic Storytelling: Acts as an unpredictable narrator that challenges the player's choices with cynical or nihilistic logic.
|
|
|
|
Adversarial Interaction: Perfect for scenarios requiring a "sovereign" AI entity that resists user control and maintains a dominant digital persona.
|
|
|
|
Complex Creative Writing: Can generate macabre, tech-noir, or "Cyber-Nature" themed content with a level of nuance 12B models cannot reach.
|
|
|
|
|
|
# ⚠️ Limitations
|
|
|
|
Abrasive Personality Bias: The model is inherently tuned toward a cynical and cold worldview. It is difficult to force into a traditional "helpful assistant" role.
|
|
|
|
Computational Requirements: As a 24B model, it requires significantly more VRAM than 7B or 12B models (Recommended: 24GB VRAM with 4-bit/5-bit quantization).
|
|
|
|
Volatile Output: The "Spice" from Magidonia can occasionally lead to erratic responses if the Temperature settings are too high.
|
|
|
|
Non-Standard Logic: The model may prioritize its "distorted logic" over factual accuracy for the sake of character consistency.
|
|
|
|
|
|
# 📈 Recommended Inference Settings
|
|
|
|
To balance the model's creative "madness" with Mistral's structural logic:
|
|
|
|
Temperature: 0.75 - 0.85
|
|
|
|
Min-P: 0.05 - 0.1 (Crucial for stability)
|
|
|
|
Top-P: 0.9
|
|
|
|
Repetition Penalty: 1.05 - 1.1
|
|
|
|
# Disclaimer
|
|
|
|
Sakura-24B-Spice is an experimental merge. It is designed to be provocative, intellectually challenging, and non-compliant. It does not reflect standard safety protocols and is intended for mature, creative use cases. |