goodfire-aiOfficial·11 Agent Skills Included

causalab

Test how language models reason with causal interventions

Tests hypotheses about how language models solve tasks internally using causal abstraction. Defines causal models, runs interchange interventions, and scores whether internal components match the proposed algorithm. Replaces ad-hoc notebooks with validated, reproducible experiment documents that any engine can execute.

108|37

↗

npx skills add goodfire-ai/causalab --all -g -y

Available:

Gives your AI agent working principles for this codebase: question approaches before implementing, find root causes, simplify code, and keep experiments clearly demonstrating their value.

All Skills in This Repository (11)

Pure Emerald Level Indicators

📦 In Repo

goodfire-ai

setup-task

Generate causalab task packages from markdown specifications with validation.

development-session

Establish development-mode context for codebase work using uv-managed environments.

plan-experiment

Create structured experiment plans with analysis DAGs and pre-flight gates.

document-issues

Append structured markdown entries to a session issues.md log.

setup-analyses

Scaffold session-local analysis code and Hydra configs from spec markdown files.

replicate-paper

Extract methodology and experiment lists from research papers to plan and run replication pipelines.

getting-started

Guide new users through Causalab codebase setup and task workflows.

research-session

Create a session directory with plan, run, and result scaffolding for research workflows.

run-experiment

Execute causal abstraction experiment plans into validated task scaffolding and runner YAML pipelines.

interpret-experiment

Generate evidence-grounded experiment reports from causalab session artifacts.

setup-methods

Scaffold session-local interpretability method stubs from set_up_method specifications.

Official

Intermediate

Frequently Asked Questions

FAQPage Schema

How to install causalab?▼

Run `npx skills add goodfire-ai/causalab --all -g -y` in your terminal to install all skills in this suite globally.

What is causal abstraction in interpretability?▼

It is a method where you write a high-level causal model of how a language model solves a task, then run interventions to test whether the model's internals actually implement that algorithm.

How do I run an intervention experiment?▼

Use the CLI with a protocol document, for example `uv run causalab run causalab/configs/protocols/interchange.json --data-root <data> --device cuda`.

Do I need a GPU to use causalab?▼

Only for running real model experiments. The onboarding demos and CPU test tiers work without any accelerator.

What experiments ship with causalab?▼

It includes ready-made protocols for activation harvesting, interchange interventions, path patching, DAS, DBM, and mean ablation, all runnable as JSON documents.