LLMs Encode Harmfulness and Refusal Separately — Reproduction

Reproducing experiments from this paper. Code is organized with configs in conf/, experiment functions in src/harm_refuse/experiments/, and an entry point at run.py.

Quick Start (uv)

• Set up environment with uv sync
• Hugging Face login for gated models/datasets: huggingface-cli login
• Run experiment: uv run run.py -cn [EXPERIMENT]

Alternative (pip)

• Create and activate a Python 3.12 virtual env.
• Install this project and deps: pip install -e .
• (Optional) huggingface-cli login
• Run: python run.py -cn [EXPERIMENT]

Notes

• Large models can be memory-heavy; use smol to validate setup.