Mini Paper Tutorials

Many researchers in interpretability write their experiments in nnsight, and you can too! Follow the applied tutorials below to recreate and experiment with results from recent published papers.

Do LMs Use Their Depth?

Skip layers to see how deeply LMs think

notebooks/mini-papers/csordas_llm_depth.ipynb

Geometry of Truth

Steer what models think is or isn't true

notebooks/mini-papers/marks_geometry_of_truth.ipynb

Function Vectors

Steer models between tasks

notebooks/mini-papers/todd_function_vectors.ipynb

Submit Your Own Ideas!

We welcome contributions to our list of applied examples. If you have an example of using nnsight in your own project, or you have an idea for a tutorial that would be helpful to you, please share it with us!

Submit Tutorial

Submit Suggestion

• Do Language Models Use Their Depth Efficiently?
• The Geometry of Truth
• Function Vectors
• Introduction
• 1️⃣ Introduction to nnsight
• 2️⃣ Task-encoding hidden states
• 3️⃣ Function Vectors
• 4️⃣ Steering Vectors in GPT2-XL
• ☆ Bonus