TransformerLens — механистическая интерпретируемость трансформеров

★ 7.7 · research

transformer-lens is a Claude Code skill that guides mechanistic interpretability research using TransformerLens — the de facto standard library for inspecting and manipulating GPT-style transformer internals via HookPoints and activation caching. Created by Neel Nanda and maintained by Bryce Meyer, TransformerLens has 2,900+ GitHub stars and supports 50+ model families including GPT-2, LLaMA, Mistral, Pythia, and Gemma. The skill covers key workflows such as activation patching and causal tracing, attention pattern analysis, circuit discovery (induction heads, IOI circuit), and direct logit attribution. It is designed for researchers reverse-engineering algorithms learned during training who need clean, programmatic access to intermediate activations and a structured approach to understanding how language models make predictions.

• #transformer-interpretability
• #mechanistic-interpretability
• #activation-analysis
• #neural-network-analysis
• #gpt-model-inspection