decoderesearch/circuit-tracer

Health v2 · maintenance only

Adoption not part of the score

2897 stars · 343 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

A Python library for circuit tracing and attribution graphs in language model interpretability, based on cross-layer MLP transcoders. It computes attribution graphs, visualizes and annotates them, and enables interventions on transcoder features.

Use cases

• trace circuits in a language model with attribution graphs
• visualize and annotate attribution graphs for LLM interpretability research
• intervene on transcoder features and observe model output changes
• study how features influence logits in transformer models
• run circuit tracing on Gemma-2 with limited GPU resources
• explore attribution graphs interactively via Neuronpedia

When to choose

• you need to compute direct effects between transcoder features, error nodes, and tokens
• you want to replicate attribution graph methods from Anthropic's circuit tracing papers
• you need feature-level interventions on a transformer model
• you want interactive visualization of interpretability graphs

When to avoid

• you need general-purpose sparse autoencoder training rather than circuit analysis
• you have no GPU access and cannot use hosted tools
• you need interpretability tooling for non-transformer architectures

Facets

library · maturity active

machine-learning deep-learning data-visualization developer-tools machine-learning deep-learning large-language-models artificial-intelligence developer-tools python cli mechanistic-interpretability attribution-graphs circuit-tracing transcoders sparse-autoencoders llm-interpretability feature-visualization gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="decoderesearch/circuit-tracer")