character-lens 0.2.0.dev0

Compute, inspect, and steer with persona/emotion vectors in language models

Character Lens

A Python library for computing, inspecting, and steering with persona/emotion vectors in language models.

Install

pip install character-lens

Or for development:

uv sync --all-extras

Overview

Character Lens extracts emotion/persona vectors from language model activations using a single-pole methodology: each trait's vector is the mean activation for that trait minus the global mean across all traits. These vectors can then be used for:

• Inspection: Project token-level activations onto persona vectors to see which concepts are active at each position
• Steering: Add persona vectors to residual stream activations during generation to influence model behavior

First-class model support: OLMo 3 7B

Quick Start

from character_lens import ActivationSteerer, extract_hidden_states

See docs/ROADMAP.md for the full development plan.

Development

# Install dependencies
uv sync --all-extras

# Run tests
uv run pytest tests/ -v --tb=short

# Run linter
uv run pre-commit run --all-files

Prior Work

The multilingual persona vectors research (v0.1.0) is archived under the v0.1.0 tag.