transformer-lens-interpretability

Activation patching on transformer models is executed by manipulating internal activations to identify causal relationships between specific components and outputs. This Skill provides scripts to trace and patch activations directly within TransformerLens.

Mechanistic interpretability in deep learning is the process of reverse-engineering neural network algorithms to understand internal circuits and attention patterns. This Skill enables detailed inspection and manipulation of transformer model components.

Yes, you can analyze induction heads and circuit behavior with TransformerLens. This Skill enables detailed circuit analysis to identify and inspect specific computational components within transformer architectures.

Yes, PyTorch is required to run circuit analysis experiments. Both TransformerLens and PyTorch are necessary dependencies for analyzing transformer architectures and executing mechanistic interpretability tasks.

A language model makes a specific prediction due to internal computational circuits processing activations through layers. This Skill helps trace those activations to pinpoint the exact transformer components responsible for a specific output.