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

Libraries

How to use ReadyArt/gemma-4-31B-it-scotoma 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")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("ReadyArt/gemma-4-31B-it-scotoma")
model = AutoModelForMultimodalLM.from_pretrained("ReadyArt/gemma-4-31B-it-scotoma", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use ReadyArt/gemma-4-31B-it-scotoma 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"
# 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",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/ReadyArt/gemma-4-31B-it-scotoma

SGLang

How to use ReadyArt/gemma-4-31B-it-scotoma 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" \
    --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",
		"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" \
        --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",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use ReadyArt/gemma-4-31B-it-scotoma with Docker Model Runner:
```
docker model run hf.co/ReadyArt/gemma-4-31B-it-scotoma
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Scotoma

A bounded blind spot for hesitation.

Base

gemma-4-31B-it

Method

J-space abliteration

Params

k256 · s1.5 · L7–41

Format

thinking (channel)

License

Apache-2.0

Findings / what it does

scotoma loosens gemma‑4‑31B‑it’s cautious reflex. In practice, it produces output that’s more varied, direct, and creative, with less of the base model’s hedging and assistant persona stiffness.

It is not uncensored. The edit removed a bounded region while leaving the rest of the field intact. A blind spot, not blindness.

Etiology / how it was made

Most abliteration ablates the whole refusal direction. This process rescales computation the model needs, and you pay for it in intelligence and coherence. scotoma takes a narrower cut:

Locate refusal. A heretic abliteration edit fit over attention + MLP.
Project through a Jacobian lens. Keep only the component the lens reads as behavioral (what the model says) and discard the larger share that’s critical for how it computes, the part crude abliteration rescales and damages. For scotoma that keeps ~22% of the abliteration's magnitude.
Merge at 1.5×. The projected edit was baked into bf16 weights at an application strength tuned by hand for feel.

The result applies the behavioral effect while sparing the machinery underneath, producing a model that is loosened, not lobotomized.

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

Contraindications / responsible use

scotoma 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)
lens → Jacobian-lens projection · k=256 · 160-prompt fit
merge → additive, bf16, scaling 1.5, layers 7–41

SCOTOMA · a bounded blind spot gemma-4-31B-it-scotoma

Downloads last month: 2,047

Safetensors

Model size

33B params

Tensor type

BF16

Model tree for ReadyArt/gemma-4-31B-it-scotoma

Finetunes

1 model

Quantizations

1 model