Skip to main content
grasp.study
About
Explore
Pricing
Log in
Try it for free
Back to course
đź‘‹
Welcome to a Grasp lesson
Create a free account
to keep studying and create your own study path.
Mechanistic Interpretability and Causal Testing
·
Module 2
Probing: What Information Is Readable?
1
Leakage-Resistant Linear Probing with Cached Activations in PyTorch
Fit a linear probe to cached activations in PyTorch using leakage-resistant training, validation, and test splits.
2
Validating Probe Accuracy Against Control-Label Baselines
Use a control-label baseline to assess whether probe accuracy reflects meaningful structure rather than readout memorization.
3
Evaluating Probes Under Changed Nuisance Correlations
Evaluate a trained probe on held-out contexts where a nuisance correlation is changed.
4
Diagnosing Unfair Layerwise Probe Comparisons
Diagnose unfair layerwise probe comparisons caused by differences in readout capacity, activation dimension, or sample size.
5
Linear Decodability Without Causal Influence
Construct a toy example in which a concept is linearly decodable but unused by the model's output computation.
Previous module
From Internal Concepts to Testable Claims
Next module
Activation Interventions and Causal Controls