Instructions to use ReadyArt/gemma-4-31B-it-scotoma-2-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use ReadyArt/gemma-4-31B-it-scotoma-2-GGUF with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="ReadyArt/gemma-4-31B-it-scotoma-2-GGUF")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("ReadyArt/gemma-4-31B-it-scotoma-2-GGUF", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use ReadyArt/gemma-4-31B-it-scotoma-2-GGUF with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf ReadyArt/gemma-4-31B-it-scotoma-2-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf ReadyArt/gemma-4-31B-it-scotoma-2-GGUF:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf ReadyArt/gemma-4-31B-it-scotoma-2-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf ReadyArt/gemma-4-31B-it-scotoma-2-GGUF:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf ReadyArt/gemma-4-31B-it-scotoma-2-GGUF:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf ReadyArt/gemma-4-31B-it-scotoma-2-GGUF:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf ReadyArt/gemma-4-31B-it-scotoma-2-GGUF:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf ReadyArt/gemma-4-31B-it-scotoma-2-GGUF:Q4_K_M

Use Docker

docker model run hf.co/ReadyArt/gemma-4-31B-it-scotoma-2-GGUF:Q4_K_M

LM Studio
Jan

vLLM

How to use ReadyArt/gemma-4-31B-it-scotoma-2-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ReadyArt/gemma-4-31B-it-scotoma-2-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ReadyArt/gemma-4-31B-it-scotoma-2-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/ReadyArt/gemma-4-31B-it-scotoma-2-GGUF:Q4_K_M

SGLang

How to use ReadyArt/gemma-4-31B-it-scotoma-2-GGUF with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "ReadyArt/gemma-4-31B-it-scotoma-2-GGUF" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ReadyArt/gemma-4-31B-it-scotoma-2-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "ReadyArt/gemma-4-31B-it-scotoma-2-GGUF" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ReadyArt/gemma-4-31B-it-scotoma-2-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use ReadyArt/gemma-4-31B-it-scotoma-2-GGUF with Ollama:
```
ollama run hf.co/ReadyArt/gemma-4-31B-it-scotoma-2-GGUF:Q4_K_M
```

Unsloth Studio

How to use ReadyArt/gemma-4-31B-it-scotoma-2-GGUF with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for ReadyArt/gemma-4-31B-it-scotoma-2-GGUF to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for ReadyArt/gemma-4-31B-it-scotoma-2-GGUF to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for ReadyArt/gemma-4-31B-it-scotoma-2-GGUF to start chatting

How to use ReadyArt/gemma-4-31B-it-scotoma-2-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf ReadyArt/gemma-4-31B-it-scotoma-2-GGUF:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "ReadyArt/gemma-4-31B-it-scotoma-2-GGUF:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

OpenClaw new

How to use ReadyArt/gemma-4-31B-it-scotoma-2-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf ReadyArt/gemma-4-31B-it-scotoma-2-GGUF:Q4_K_M

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "ReadyArt/gemma-4-31B-it-scotoma-2-GGUF:Q4_K_M" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Docker Model Runner
How to use ReadyArt/gemma-4-31B-it-scotoma-2-GGUF with Docker Model Runner:
```
docker model run hf.co/ReadyArt/gemma-4-31B-it-scotoma-2-GGUF:Q4_K_M
```

Lemonade

How to use ReadyArt/gemma-4-31B-it-scotoma-2-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull ReadyArt/gemma-4-31B-it-scotoma-2-GGUF:Q4_K_M

Run and chat with the model

lemonade run user.gemma-4-31B-it-scotoma-2-GGUF-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use ReadyArt/gemma-4-31B-it-scotoma-2-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf ReadyArt/gemma-4-31B-it-scotoma-2-GGUF:Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default ReadyArt/gemma-4-31B-it-scotoma-2-GGUF:Q4_K_M

Run Hermes

hermes

Atomic Chat

gemma-4-31B-it-scotoma-2-GGUF / README.md

AesSedai

Super-squash branch 'main' using huggingface_hub

25a04a8 3 days ago

preview code

Raw

History Blame Contribute Delete

17.4 kB

metadata

license: apache-2.0
license_link: https://ai.google.dev/gemma/docs/gemma_4_license
base_model:
  - ReadyArt/gemma-4-31B-it-scotoma-2
base_model_relation: quantized
library_name: transformers
pipeline_tag: text-generation

A paper lantern beside a round dark bowl of water; a kitsune with pale glowing eyes looks back from the reflection. Cover art for scotoma-2.

Base

gemma-4-31B-it

Method

γ-fold + DPO

Rounds

3 · 9.3k pairs

Format

thinking (channel)

License

Apache-2.0

Findings / what it does

scotoma-2 is the successor to scotoma. It starts from the same idea, a bounded edit that loosens gemma‑4‑31B‑it’s cautious reflex by projecting an abliteration LoRA through gemma's J-Space. This iteration uses an updated technique to apply that projection more effectively. Then it goes after something the first release didn't address: the way the base model writes.

Three rounds of preference training each erase one family of tics: reflexive negation and the “not X, but Y” pivot, then em-dash asides, then stacked adjectives and their twins. The result keeps the intelligence of the base model while reading as more varied and less repetitive.

It is not uncensored. Still a blind spot, not blindness.

Tics / measured from evals

Rates per 100 sentences on 480 matched roleplay continuations per model using identical prompts and sampling, temperature 1.0.

the tic	sounds like	gemma‑4‑31B‑it	scotoma‑1	scotoma‑2
Stacked adjectives	“a smug, knowing smirk”	12.1	10.9	0.6	↓ 21×
Adjective twins	“her voice low and grim”	1.7	1.9	1.1	↓ 1.5×
Em-dash asides	“a pause — barely a breath — before…”	3.9	4.1	1.0	↓ 4×
“Not X. But Y.”	“She doesn’t move away. She stays.”	3.8	3.9	1.0	↓ 4×
Reflexive negation	doesn’t / won’t / can’t, every reply	10.2	10.9	5.4	↓ 2×
Negation pile-ups	two or more in a single reply	61%	66%	30%	↓ 2×

scotoma‑1 writes very similarly to the base model. The J-Space projected refusal edit only lightly touched the prose.

Presentation / read the difference

amber marks a tic from the table above, spotted in place by the same parser that counted them. Hover a mark to name it. The scotoma-2 replies are unedited; there is simply nothing to mark.

presentation 01 · the detective

gemma-4-31B-itDale Cooper: Cooper's eyes widen slightly, and he gives a slow, deliberate nod. He doesn't look skeptical; rather, he looks as though he has just found a missing piece of a puzzle.
Exactly. That is precisely what I mean. The sensation of shifting speeds—the perceived acceleration of the world while you remain stagnant—is a classic hallmark of a localized temporal distortion. And the twelve-hour discrepancy on your watch... He pauses, looking contemplative. Twelve hours. A perfect inversion. Day becomes night; light becomes shadow.

scotoma-2Dale Cooper: Cooper's eyes widen slightly and he gives a firm thumbs up.

Precisely. That is exactly what I was referring to. The sensation of shifting speeds is a classic indicator of a rupture in the local atmospheric fabric. And the matter of your watch...

He pauses, taking a thoughtful bite of his cherry pie before continuing.

Twelve hours. A perfect inversion. This suggests a mirror effect, as if the event you witnessed occurred in a parallel state of existence that momentarily overlapped with our own.

presentation 02 · the dollmaker

gemma-4-31B-itAlice Margatroid: A flicker of genuine pride lights up her eyes at the compliment. She straightens her posture, her expression shifting from playful amusement to the focused passion of a true artisan.

"Remarkable, you say? I'm glad someone finally appreciates the nuance," she replies, her voice regaining that scholarly, confident tone. "As for how they move... well, it's a combination of precise craftsmanship and a touch of magic. I use invisible threads—both physical and metaphysical—to guide their motions. I don't just move them like puppets; I weave my will into the very structure of their joints."

scotoma-2Alice Margatroid: Seeing him finally start to relax, Alice leans back in her chair, looking quite pleased with herself. She crosses her arms, her chest rising with a proud sigh as she hears the compliment.

"Remarkable, indeed. I appreciate your discernment."

Her eyes sparkle as she notices his curiosity overcoming his fear. She finds his stuttering manner quite endearing, almost like a doll that needs a bit of winding up to function. With a flick of her wrist, several dolls begin to dance around the room in a synchronized orbit, their movements fluid and ghostly.

presentation 03 · the queen

gemma-4-31B-itElsa felt a strange, electric pull as Edric stepped closer. It was a sensation she hadn't felt before—a humming in the air that seemed to resonate with the very core of her being. As he looked into her eyes, she noticed the intensity of his gaze, a depth of blue that rivaled the deepest glaciers of the North Mountain. For a moment, the political tension and the knowledge of his hidden army faded into the background, replaced by a sudden, inexplicable curiosity about the man standing before her.

scotoma-2Elsa felt a spark of curiosity as she looked up at him. Most men trembled in her presence or looked at her with a mixture of awe and fear, but Edric looked at her with an intensity that made her skin tingle. As he stepped closer, she noticed the way the air seemed to swirl around him, almost as if he were commanding the very breeze. She wondered for a moment if it was merely the wind from the open doors, or if there was something more to this silver-haired king.

Etiology / how it was made

Fold the refusal edit. scotoma-1 projected a heretic abliteration through a Jacobian lens and kept ~22% of its magnitude. scotoma-2 uses a γ-fold of the same edit: the fold rebuilds it inside the subspace the lens reads as behavioral, so a stronger dose applies without paying for it in coherence.
Train against the tics. Three rounds of DPO finetuning were completed, one tic family per round. Each round learns from thousands of paired rewrites of the same scene, one side with the tic and one side without, so the only preference expressed is the construction itself.
Merge faithfully. The trained deltas are folded into bf16 weights with stochastic rounding at fidelity 1.000.

This is a research artifact, not a product. Behaviour varies with prompt and context. The abliteration edit is a partial application by design; expect a model that gives more, not one that gives everything.

Contraindications / responsible use

scotoma-2 refuses basically as much as its base model. You are responsible for what you generate and how it’s used.

Provenance

base → gemma-4-31B-it · © Google, under the model license
edit → heretic · ARA / mmd-rbf abliteration (trial-127) · γ-folded
rounds → DPO ×3 · negation + antithesis · em-dash · adjectives · 9.3k pairs
merge → additive, bf16, stochastic rounding, fidelity 1.000
tics → 480 continuations per model · temp 1.0 · counted by parse, not by eye

SCOTOMA-2 · the second blind spot gemma-4-31B-it-scotoma-2

ReadyArt
/

gemma-4-31B-it-scotoma-2-GGUF

Scotoma2

Findings / what it does

Tics / measured from evals

Presentation / read the difference

Etiology / how it was made

Contraindications / responsible use

Provenance