Main Tutorials

Get Started

Predict

Control

Understand

Welcome! Follow the tutorials below to learn how to conduct interpretability research with nnsight.

Walkthrough

Learn the basics

notebooks/tutorials/get_started/walkthrough.ipynb

Access LLMs

Use our hosted models

notebooks/tutorials/get_started/start_remote_access.ipynb

Chat Templates

Format instructions with templates

notebooks/tutorials/get_started/chat_templates.ipynb

Logit Lens

Decode activations

notebooks/tutorials/probing/logit_lens.ipynb

Diffusion Lens

Explore diffusion model text embedding

notebooks/tutorials/probing/diffusion_lens.ipynb

Dictionary Learning

Sparse autoencoders

notebooks/tutorials/steering/dict_learning.ipynb

LoRA

Lightweight model adapter

notebooks/tutorials/steering/LoRA_tutorial.ipynb

Activation Patching

Causal intervention

notebooks/tutorials/causal_mediation_analysis/activation_patching.ipynb

Attribution Patching

Approximate patching

notebooks/tutorials/causal_mediation_analysis/attribution_patching.ipynb

DAS

Localizing causal variables

notebooks/tutorials/causal_mediation_analysis/DAS.ipynb

Get Started

• Walkthrough
• 1 First, let’s start small
• 2️ Bigger
• 3 I thought you said huge models?
• Next Steps
• Access LLMs with NDIF and NNsight
• Install NNsight
• Sign up for NDIF remote model access
• Choose a Model
• Access model internals
• Alter model internals
• Next steps: Run your own experiment with NDIF and NNsight
• Chat Templates
• Setup
• Applying a Chat Template
• Chat Template Parameters
• Multiple Templates
• Model Training

Predict

• Logit Lens
• Diffusion Lens

Control

• Dictionary Learning
• Setup
• Apply SAE
• LoRA for Sentiment Analysis
• Setup
• Prepare Data
• Prepare our Model
• LLM Fine Tuning

Understand

• Activation Patching
• Setup
• Patching Experiment
• Limitations
• Trying on a bigger model
• Attribution Patching
• Remote Attribution Patching
• DAS