Schism

Reshape how AI thinks, one slider at a time.

Schism is a local activation steering tool for LLMs. It lets you control model personality — creativity, humor, formality, confidence, and more — using visual sliders or CLI flags. No fine-tuning, no cloud, no ML knowledge required.

It works by injecting steering vectors into a model's hidden states at inference time, using either Sparse Autoencoders (SAE) or contrastive activation extraction.

Quick Start

pip install schism

# Download a model (~5GB)
schism download gemma-2-2b

# Launch the web UI
schism

Open http://localhost:6660, drag the sliders, type a prompt, and watch the model's personality shift in real time.

CLI Usage

# Generate with personality sliders
schism steer "Explain quantum physics" --humor 0.9 --creativity 0.8 --formality -0.6

# Use a preset
schism steer "Tell me about the ocean" --preset pirate

# Side-by-side comparison: vanilla vs steered
schism compare "What is love?" --preset sarcastic

# Pre-compute steering vectors for faster first generation
schism warmup gemma-2-2b

Steering Dimensions

Each slider maps to a steering vector computed from contrastive prompt pairs:

Slider	Description
creativity	Imaginative and expressive vs plain and literal
formality	Professional and academic vs casual and relaxed
humor	Witty and funny vs serious and dry
confidence	Assertive and decisive vs hesitant and uncertain
verbosity	Detailed and thorough vs brief and concise
empathy	Warm and compassionate vs cold and analytical
technical	Expert jargon vs simple language

All sliders range from -1.0 to +1.0. Create custom features with schism features create.

Presets

11 built-in personality presets:

Creative · Formal · Sarcastic · Concise · Pirate · ELI5 · Academic · Noir Detective · Zen Master · Chaos · Robot

# List all presets
schism presets list

# Show preset details
schism presets show pirate

# Export for sharing
schism presets export pirate -o pirate.json

Supported Models

Model	HuggingFace ID	SAE Support
gemma-2-2b	google/gemma-2-2b	Yes (Gemma Scope)
gemma-2-2b-it	google/gemma-2-2b-it	Yes (Gemma Scope)
llama-3.2-3b	meta-llama/Llama-3.2-3B	No (contrastive)
llama-3.2-3b-instruct	meta-llama/Llama-3.2-3B-Instruct	No (contrastive)

Models with SAE support use Sparse Autoencoders for more interpretable steering. Models without SAE fall back to contrastive activation extraction, which still works well.

How It Works

1. Contrastive prompts define each steering dimension (e.g. "be creative" vs "be plain")
2. Activation extraction captures the model's internal representations for each set of prompts
3. Steering vectors are computed as the difference between positive and negative activations
4. At inference time, these vectors are injected into the model's hidden states at a target layer, scaled by the slider value

Based on research from Arditi et al. 2024 and FGAA (2025).

Docker

docker build -t schism .
docker run -p 6660:6660 schism

For GPU support:

docker run --gpus all -p 6660:6660 schism

Development

git clone https://github.com/suislanchez/schism.git
cd schism
pip install -e ".[test]"
pytest tests/ -v

License

MIT