transformer-lens-interpretability

Activation patching on a GPT-style transformer requires caching internal activations and using HookPoints to intervene. You can compare clean versus corrupted prompts by patching cached residual streams or attention patterns to test causal tracing.

To find induction heads, you cache attention patterns across layers and analyze the model's activations. This skill provides workflows to map influential heads and inspect attention circuits for sequence copying behavior.

Yes, logit lens inspection allows you to trace transformer predictions by applying the unembedding matrix to hidden residual stream states at each layer. This reveals how specific layers progressively affect the final output logits.

Circuit analysis without manual instrumentation involves using TransformerLens to expose and edit internal activations via hooks. You can perform ablation and steering experiments to map influential heads and test layer effects on predictions.

Steering experiments on cached activations use hook-based interventions to modify residual streams during the forward pass. This allows you to alter the model's behavior and test how specific layers or attention heads affect predictions.

Analyzing transformer model internals for mechanistic interpretability involves using TransformerLens to cache activations and apply hook-based interventions. This exposes residual streams and attention patterns to map circuit behavior.