Bias 🎯

Steer Language Models with Interpretable SAE Features

Bias is a Python library for steering LLM behavior using Sparse Autoencoder (SAE) features from Neuronpedia. Instead of prompt engineering or fine-tuning, simply describe the behavior you want.

---

Installation

# From GitHub
pip install git+https://github.com/codewithdark-git/bias.git

# With dev tools
pip install "bias[dev] @ git+https://github.com/codewithdark-git/bias.git"

Requirements: Python 3.11+, PyTorch 2.5+

---

Quick Start

Basic Usage

from bias import Bias

# Initialize
bias = Bias("gpt2")

# Steer toward a concept
bias.steer("professional formal writing", intensity=2.0)

# Generate
output = bias.generate("Write an email about the project:")
print(output)

# Reset
bias.reset()

With Configuration (Recommended)

from bias import Bias, BiasConfig

# Create configuration with your API key
config = BiasConfig(
    api_key="your-neuronpedia-api-key",  # Get from neuronpedia.org
    model="gpt2",
    device="auto",
)

# Pass config to Bias
bias = Bias(config=config)
bias.steer("professional writing")
output = bias.generate("Write an email:")

Using Environment Variables

# Set your API key as environment variable
export NEURONPEDIA_API_KEY="your-api-key"

from bias import Bias, BiasConfig

# Config automatically reads from environment
config = BiasConfig(model="gpt2")
bias = Bias(config=config)

# Or use from_env() for full environment configuration
config = BiasConfig.from_env()
bias = Bias(config=config)

---

Configuration

BiasConfig Options

Parameter	Type	Default	Description
api_key	str	None	Neuronpedia API key (or use NEURONPEDIA_API_KEY env var)
model	str	"gpt2"	Model name ("gpt2", "gpt2-medium", etc.)
layer	int	auto	Steering layer (auto-selected per model)
device	str	"auto"	Device ("auto", "cuda", "cpu", "mps")
dtype	str	"float16"	Precision ("float16", "bfloat16", "float32")
intensity	float	1.0	Default steering intensity
num_features	int	5	Features per concept

Full Configuration Example

from bias import Bias, BiasConfig

config = BiasConfig(
    # API Settings
    api_key="your-api-key",        # Neuronpedia API key
    
    # Model Settings
    model="gpt2-medium",           # Model to use
    layer=12,                      # Steering layer
    sae_id="res-jb",              # SAE variant
    
    # Device Settings
    device="cuda",                 # GPU acceleration
    dtype="float16",               # Half precision
    
    # Steering Defaults
    intensity=1.5,                 # Default intensity
    num_features=5,                # Features per concept
)

bias = Bias(config=config)

Quick Configure Function

from bias import configure, Bias

# One-liner configuration
config = configure(api_key="your-key", model="gpt2-medium")
bias = Bias(config=config)

---

Method Chaining

output = (
    Bias("gpt2")
    .steer("creative poetic", intensity=2.0)
    .generate("The moonlight danced upon")
)

Compare Outputs

bias.steer("formal academic", intensity=3.0)
results = bias.compare("Explain gravity:")
print("Unsteered:", results['unsteered'])
print("Steered:", results['steered'])

---

CLI

# Generate with steering
bias generate "Write a poem:" -c "romantic" -i 2.0

# Discover features
bias discover "technical language"

# Interactive mode
bias interactive

---

Core API

Method	Description
Bias(config=config)	Initialize with configuration object
Bias(model, api_key=...)	Initialize with parameters
steer(concept, intensity)	Steer toward a concept
generate(prompt)	Generate text
compare(prompt)	Compare steered vs unsteered
discover(concept)	Find features for a concept
reset()	Clear steering

---

Supported Models

Model	Recommended Layer	Neuronpedia ID
gpt2	6	gpt2-small
gpt2-medium	12	gpt2-medium
gpt2-large	18	gpt2-large
gpt2-xl	24	gpt2-xl

---

Environment Variables

Variable	Description
NEURONPEDIA_API_KEY	Your Neuronpedia API key
BIAS_MODEL	Default model name
BIAS_LAYER	Default layer number
BIAS_DEVICE	Default device

---

How It Works

Bias uses Sparse Autoencoder (SAE) features from Neuronpedia to steer models. Each feature represents an interpretable concept (formality, sentiment, etc.). Adding these feature vectors to model activations shifts behavior toward that concept.

📖 Full Documentation — Detailed guides on steering, SAEs, and the Neuronpedia integration.

---

License

MIT License — see LICENSE

---

Made with 🎯 by codewithdark-git