Stefan Heimersheim Stefan-Heimersheim

Mechanistic Interpretability researcher @ Apollo Research

Stefan-Heimersheim / runpod_setup.sh

Last active September 30, 2025 14:31

	# Make sure to add your SSH key to your RunPod account
	echo "export HF_HOME=/workspace/hf_home/" >> ~/.bashrc
	apt-get update
	apt-get upgrade -y
	apt-get install -y sudo git vim ssh net-tools htop curl zip unzip tmux rsync libopenmpi-dev iputils-ping make fzf restic ripgrep wget pandoc poppler-utils pigz bzip2 nano locales
	pip install uv
	# Alternative: curl -LsSf https://astral.sh/uv/install.sh \| sh
	uv pip install --system --compile-bytecode ipykernel kaleido nbformat numpy scipy scikit-learn scikit-image transformers datasets torchvision pandas matplotlib seaborn plotly jaxtyping einops tqdm ruff basedpyright umap-learn ipywidgets virtualenv pytest git+https://github.com/callummcdougall/eindex.git transformer_lens nnsight
	apt-get install -y libnss3 libatk-bridge2.0-0 libcups2 libxcomposite1 libxdamage1 libxfixes3 libxrandr2 libgbm1 libxkbcommon0 libpango-1.0-0 libcairo2 libasound2
	# Optional:

Stefan-Heimersheim / figure1.py

Last active August 18, 2025 11:27

Code: [Interim research report] Activation plateaus & sensitive directions in GPT2

	# 1.py (Figure 1 is a zoom into one panel of Figure 3/4)
	import os
	import random
	from collections import defaultdict
	from collections.abc import Callable
	from dataclasses import dataclass

	import matplotlib
	import matplotlib.pyplot as plt

Stefan-Heimersheim / tms_spiky_activation_space.py

Created April 2, 2025 10:38

Demo of activation space geometry in a simple Toy Model of Superposition setup

	import itertools

	import matplotlib as mpl
	import matplotlib.pyplot as plt
	import numpy as np
	import plotly.graph_objects as go
	import torch
	import torch.nn as nn
	import torch.optim as optim
	from jaxtyping import Float

Stefan-Heimersheim / clustering_pythia.py

Last active February 11, 2025 11:26

	import math
	import os
	from abc import abstractmethod
	from collections import defaultdict
	from functools import partial
	from typing import Optional

	import joblib
	import numpy as np
	import sae_bench.custom_saes.base_sae as base_sae