domains / families / science / transformer-lens-interpretability

transformer-lens-interpretability

separate

Provides guidance for mechanistic interpretability research using TransformerLens to inspect and manipulate transformer internals via HookPoints and activation caching. Use when reverse-engineering model algorithms, studying attention patterns, or performing activation patching experiments.

When to use

• Reverse-engineer algorithms learned during training
• Perform activation patching / causal tracing experiments
• Study attention patterns and information flow
• Analyze circuits (e.g., induction heads, IOI circuit)
• Cache and inspect intermediate activations
• Apply direct logit attribution
• You need to work with non-transformer architectures → Use **nnsight** or **pyvene**
• You want to train/analyze Sparse Autoencoders → Use **SAELens**

When not to use

• Work with non-transformer architectures
• Train/analyze Sparse Autoencoders
• Remote execution on massive models
• Higher-level causal intervention abstractions

Skill page

Source