TransformerLensOrg/TransformerLens

Health v2 · maintenance only

Adoption not part of the score

3949 stars · 709 forks observed · 2026-10-08

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

TransformerLens is a Python library for mechanistic interpretability of GPT-style transformer language models. It lets researchers load thousands of open-source models and inspect, cache, and edit internal activations via hooks as the model runs.

Use cases

• inspect internal activations of transformer models
• reverse engineer algorithms learned by language models
• run activation patching experiments on GPT-2 style models
• hook and edit model activations during a forward pass
• study attention heads and circuits in LLMs
• load HuggingFace models for interpretability research

When to choose

• you want to analyze or edit internal activations of open-source language models
• you're doing mechanistic interpretability research on transformers
• you need hook-based access to model internals across many architectures

When to avoid

• you just need to run inference or fine-tune models without inspecting internals
• you need a production serving or deployment tool
• you work with non-transformer model architectures

Facets

library · maturity active

machine-learning llm-inference developer-tools sdk machine-learning deep-learning large-language-models artificial-intelligence python cross-platform mechanistic-interpretability activation-caching hooking transformers explainability pytorch research gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="TransformerLensOrg/TransformerLens")