66 lines
2.9 KiB
Markdown
66 lines
2.9 KiB
Markdown
|
---
|
|||
|
|
license: apache-2.0
|
|||
|
|
language:
|
|||
|
|
- en
|
|||
|
|
base_model:
|
|||
|
|
- Vortex5/Silver-Siren-12B
|
|||
|
|
pipeline_tag: text-generation
|
|||
|
|
tags:
|
|||
|
|
- mistral
|
|||
|
|
- sft
|
|||
|
|
- unsloth
|
|||
|
|
- style-tune
|
|||
|
|
- roleplay
|
|||
|
|
- creative
|
|||
|
|
- 12b
|
|||
|
|
- nemo
|
|||
|
|
datasets:
|
|||
|
|
- ewald1976/opus48-gpt55-asimov.jsonl
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
|
|||
|
|
# Silver-Siren-ST-12B (Mistral-Nemo-12B-based StyleTune)
|
|||
|
|
|
|||
|
|
A while ago, I stumbled upon Gryphe's report on "Style-Tuning" (`Gemma-4-31B-StyleTune`) by chance. At least for me, this concept was entirely new, but it sounded incredibly promising—so I just had to test it out myself.
|
|||
|
|
|
|||
|
|
This model is the result of that experiment: A technical test in **surgical style-tuning** to verify if isolating style changes to a single tensor (`lm_head`) while freezing the core weights works effectively on the **Mistral-NeMo-12B** architecture.
|
|||
|
|
|
|||
|
|
|
|||
|
|
## Methodology & Concept
|
|||
|
|
|
|||
|
|
Normally, a fine-tune alters as many layers as possible to align both the reasoning and formatting of a model to a new dataset. This "StyleTune" approach does the exact opposite:
|
|||
|
|
1. **Freeze everything:** All attention and MLP layers (Layers 0–39) remain completely untouched. The underlying logic, world knowledge, and instruction-following capabilities are preserved exactly as they were.
|
|||
|
|
2. **Target the Language Center:** Only **one single tensor**—the `lm_head` (output projection)—is trained.
|
|||
|
|
|
|||
|
|
By retraining only the `lm_head`, the model doesn't become "smarter" or "dumber," but its vocabulary, sentence structure, and prose quality are completely recalibrated. It changes the *voice* of the model, not its brain.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## The Target Base: Why this model?
|
|||
|
|
|
|||
|
|
To demonstrate the contrast and efficacy of this method, I deliberately chose **Vortex5/Silver-Siren-12B** as the target base.
|
|||
|
|
|
|||
|
|
> **Important Note:** This choice is purely technical and meant with the utmost respect for the original author. `Silver-Siren-12B` is a highly popular, emotion-forward merge (incorporating models like *Dark-Nexus*, *Elysian-Sunrise*, *LunaMaid*, and others) that is highly optimized for sensational, dramatic, and immersive interactions. Because its native style profile is so distinct, it served as the perfect benchmark to test if a pure `lm_head` tune could cleanly overwrite a deeply baked-in stylistic bias without degrading the underlying merge quality.
|
|||
|
|
|
|||
|
|
By exposing this base to a highly curated, classical sci-fi literary dataset (inspired by Asimov, Huxley, and Lem), the model underwent a dramatic transformation in its prose delivery.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Training Details & Parameters
|
|||
|
|
|
|||
|
|
* **Epochs:** 3
|
|||
|
|
* **Learning Rate:** 4e-4 (Linear Scheduler)
|
|||
|
|
* **Target Modules:** `lm_head` *only* (all other linear layers frozen)
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Recommended Sampler Settings
|
|||
|
|
* **Temperature:** 0.7 - 0.9
|
|||
|
|
* **Min_P:** 0.05
|
|||
|
|
* **Top_P:** 0.95
|
|||
|
|
* **Repetition Penalty:** 1.05
|
|||
|
|
|
|||
|
|
## Thank you
|
|||
|
|
|
|||
|
|
- to Gryphe for posting this excellent finding. https://huggingface.co/Gryphe/Gemma-4-31B-StyleTune
|
|||
|
|
- to Vortex5 for creating this model.
|