Modelpublic

Wiself/Dark-Scarlett-v1.0-26B-A4B-Voice

sourceHugging Faceapache-2.0updated 19d agoView on Hugging Face

1likes728downloads

Model Card

The Dark Scarlett voice for Gemma 4 26B MoE — as a download.

Dark-Scarlett-v1.0-26B-A4B-Voice

Dark Scarlett seems like a cool model. Try its voice on your GGUF, see how it sounds — nothing crazy.

This is the voice of ReadyArt/Dark-Scarlett-v1.0-26B-A4B in a single portable tensor (~1.5 GB). If you already have a Gemma 4 26B A4B GGUF, cast it on with the Voice tool and hear how it reads — your base underneath, Scarlett's writing on top.

What it is

Dark Scarlett is an uncensored M→F roleplay LoRA (rank 32, 2 epochs, 12,211 prompts) on the Gemma 4 26B instruct base. We lift the output head — the part that shapes the writing — and skip the rest, so the download is one tensor instead of a full model.

Two steps

bash

# 1. Get the voice tool (one-time): https://huggingface.co/Wiself/voice
python3 voice.py path   # optional, puts `voice` on your PATH

# 2. Cast it onto any Gemma 4 26B A4B GGUF you already have
voice cast ./gemma-4-26b-a4b-it-Q4_K_M.gguf voice.safetensors --out ./voiced/gemma-4-26b-a4b-scarlett.gguf

Run it:

bash

llama serve -m ./voiced/gemma-4-26b-a4b-scarlett.gguf --jinja

One file out. No adapter at runtime, no LoRA, no second model.

The numbers

Our own probe, not the source author's benchmarks — one prompt, five builds, single greedy draws each (first probe, more coming; full data as tensorwars.md in the Voice tool repo). Cast onto Orion (TheDrummer_Orion-26B-A4B-v1):

—127 voice-only 4-grams — phrases the voiced builds share with each other and never with the base. Longest shared spans reach 8 words (e.g. the clatter of cart wheels on basalt paving).
—Donor fidelity — the output-slot build shares 4 exact phrases with Scarlett herself (distant roar of the crowd, the grain dole, the emperor himself, basalt paving stones). Trigram overlap with base sits at 0.026–0.036.
—Slot matters — same voice lands differently per slot (output-only vs embed-only share just 0.058 trigram overlap). Default cast speaks through output.weight; --speak forces it.

Works on more than the base model

Because the voice is just a tensor, it casts onto any compatible Gemma 4 26B A4B GGUF:

—Different finetunes — keep the stronger base, gain Scarlett's voice
—Quantized GGUFs — any quant; your tensors stay untouched, only the head is replaced

<details> <summary>Loops on abliterated targets?</summary>

On heavily modified targets a direct cast can loop. The delta path fixes it — it carries only the difference from the base instruct, so MoE routing stays calm:

bash

voice delta voice.safetensors --base google/gemma-4-26B-A4B-it
voice cast ./model.gguf delta-voice.safetensors --out ./voiced/model.gguf

</details>

What's inside

—voice.safetensors — the lm_head.weight tensor, source dtype preserved, arch-derived shape [262144, 2816], ~1.5 GB
—voice.json — metadata: source, dtype, shape, architecture

Confirm after download: voice info voice.safetensors → lm_head.weight · [262144, 2816].

Compatibility

Target	Works?
Any Gemma 4 26B A4B GGUF (any quant)	✅
QAT-derived GGUFs	✅ works, but the QAT voice is the exact match
Gemma 4 other sizes (9B, 12B, 31B)	❌ shape mismatch — use the matching voice
Non-Gemma architectures	❌

Notes

—Sampler tips from the source card: Top-P 0.92, temp 1.0, no frequency/presence penalty. Thinking models: enable_thinking: true, medium reasoning effort.
—Usage terms carry over: personal use only, 18+, you accept full responsibility for outputs — see the source card before sharing voiced models.
—Casting quantizes only the head to Q8_0 (near-lossless); all other tensors byte-copied.

Credits

ReadyArt — Dark Scarlett finetune (dataset generation: GECFDO; fine-tuning: FrenzyBiscuit). The Voice tool lifts and casts it. Base model is Google's, under Gemma terms.