Interpretability Beginner

环境

conda create -n interpretability python=3.10
conda activate interpretability
pip install -r requirements.txt

• requirements.txt — lec1–lec12 所需（Chapter 1 可解释性）
• requirements-rl.txt — 可选，ARENA Chapter 2（RL），需 Python 3.11+
• requirements-evals.txt — 可选，ARENA Chapter 3（evals）

在 VS Code / Jupyter 中选择 interpretability kernel。Notebook 的 Setup cell 不会再自动 pip install，只会校验依赖是否已安装。

每节目录包含：

• lecture.md — 讲义
• lecX_exercises.ipynb — 练习 notebook
• lecX_solutions.ipynb — 参考答案
• 辅助代码（tests.py、solutions.py 等）

目录

目录	原编号	主题
lec1_transformer_from_scratch	1.1 Transformer from Scratch	transformer from scratch
lec2_intro_to_mech_interp	1.2 Intro to Mech Interp	intro to mech interp
lec3_linear_probes	1.3.1 Linear Probes	linear probes
lec4_function_vectors_and_model_steering	1.3.2 Function Vectors & Model Steering	function vectors and model steering
lec5_interp_with_saes	1.3.3 Interpretability with SAEs	interp with saes
lec6_activation_oracles	1.3.4 Activation Oracles	activation oracles
lec7_indirect_object_identification	1.4.1 Indirect Object Identification	indirect object identification
lec8_sae_circuits	1.4.2 SAE Circuits	sae circuits
lec9_balanced_bracket_classifier	1.5.1 Balanced Bracket Classifier	balanced bracket classifier
lec10_grokking_and_modular_arithmetic	1.5.2 Grokking & Modular Arithmetic	grokking and modular arithmetic
lec11_othellogpt	1.5.3 OthelloGPT	othellogpt
lec12_toy_models_of_superposition_and_saes	1.5.4 Toy Models of Superposition & SAEs	toy models of superposition and saes
lec13_monthly_algorithmic_problems	1.5.X Monthly Algorithmic Problems	monthly algorithmic problems