Set up a small open model and TransformerLens-compatible…

1. 01

   Concepts & system design

   60 min· read · map the mechanismtimer

   R34 — TransformerLens Getting Started, HookPoints & ActivationCache

   Read the assigned sections; map the mechanism, threat model and trade-offs.

2. 02

   Project implementation

   120 min· build · integrate · debugtimer

   Set up a local TransformerLens environment with a small open-weight model (e.g., GPT-2 Small or Pythia/Qwen small). Inspect transformer blocks, attention heads, MLP layers, the residual stream, logits, and activation caches.

   Build, integrate, debug and improve the increment in the active project.

3. 03

   Security lab

   60 min· attack your own systemtimer

   Trace one prompt's token-by-token residual stream activations locally and verify deterministic activation caching.

   Test the abuse case, failure path, authorization bypass or resource limit.

4. 04

   Verification & evidence

   30 min· test · record evidencetimer

   Reproducible local TransformerLens setup and activation-caching notebook.

   Add or run regression tests; record the 9-point evidence record.

5. 05

   Deliberate practice & Socratic review

   30 min· explain aloud · update gapstimer

   Answer the 8 Socratic questions without notes: What? Why? How? What if? Why not? How can I prove it? Where can it break? What have I learned?

   Re-explain the hardest concept aloud, inspect one edge case, update the gap log.

6. total5 h

• R34TransformerLens Docs↗ transformerlensorg.github.io

  HookedTransformer, residual stream, ActivationCache, Logit Lens, and Activation Patching tutorials.

Read only the named sections needed for today's task. Reading sits inside the 60-minute concepts block; it is not extra time.

• docsTransformerLens · reference
  Getting Started - TransformerLens Documentation

  Covers installing TransformerLens, loading a HookedTransformer and running a first forward pass with cache, the exact setup path for inspecting blocks, heads, MLPs and the residual stream locally.

  R34checked 2026-11-11
• videoJ Rosser · 12 min
  Speedrunning a Mech Interp Research Setup (Remote GPU, Torch, TransformerLens, Cuda, SSH, VS Code)

  Shows a reproducible environment build with Torch, CUDA and TransformerLens on a remote GPU, so you can stand up the small-model workspace without dependency drift.

  R34checked 2026-11-11
• videoEuroPython Conference · 30 min
  Hacking LLMs: An Introduction to Mechanistic Interpretability — Jenny Vega

  Walks the residual stream, attention heads and MLP layers with code, giving you the mental model for reading HookedTransformer's named hook points before you start poking at GPT-2 Small.

  R34checked 2026-11-11
• articleNeel Nanda · ~25 min read
  How To Become A Mechanistic Interpretability Researcher

  Explains what to actually inspect in a small model and how to build the habit of forming and testing hypotheses about components, which is the working method for this setup day.

  R34moved 2026-11-11
• repoGitHub (TransformerLensOrg) · library + demos
  GitHub - TransformerLensOrg/TransformerLens: A library for mechanistic interpretability of GPT-style language models

  Ships the library, demo notebooks and hook documentation you install to enumerate modules, run_with_cache, and read logits from the residual stream on a small open-weight model.

  R34checked 2026-11-11

A lab is complete only when it produces evidence. Installing PyRIT, Garak, Presidio or TransformerLens is not completion. Saved on this device only.