Neural Multimodal Transformer Mechanistic Interpretability

Understanding how neural networks process information is critical for building trustworthy AI systems. This package provides hands-on tools for inspecting model activations, discovering circuits, and performing causal interventions.

• TransformerLens-based interpretability toolkit
• Path patching implementation for circuit discovery
• Activation patching and causal intervention utilities
• Probing classifiers for feature analysis
• Visualization tools for attention patterns and activations

• Introduction to Mechanistic Interpretability
• Path Patching Tutorial
• Circuit Discovery with ACDC
• Probing Classifier Walkthrough
• Attention Pattern Visualization