skills.rest
goodfire-ai
goodfire-aiOFFICIAL
Explore all skills by this author →
Website𝕏@goodfireaiGitHub

run-experiment

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

108|37|Updated Apr 25, 2025
Search Tags:#intervention evaluation#mechanistic interpretability#experiment orchestration#causal abstraction#token alignment#language model experiments#hydra runner configs
One-click install
npx skills add https://github.com/goodfire-ai/causalab --skill run-experiment-goodfire-ai
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: run-experiment
Source: https://github.com/goodfire-ai/causalab/tree/main/.claude/skills/run-experiment
Command: npx skills add https://github.com/goodfire-ai/causalab --skill run-experiment-goodfire-ai

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

It helps you execute the full causal abstraction experiment workflow—turning a session plan into validated task setup, runner configuration, execution, and artifact verification—so results are produced without manual orchestration.

Core Features & Use Cases

  • End-to-end execution from a plan: Consumes the outputs of /plan-experiment (RESEARCH_OBJECTIVE.md + PLAN.md) and runs the specified experiment chain.
  • Autonomous setup with validation: Optionally creates missing tasks via /setup-task, then runs task tests to catch forward-pass, token-position, and token-alignment issues before execution.
  • Runner-config driven runs: Builds and snapshots Hydra runner YAML configs in the session directory, runs debug passes, and then executes the full preset via scripts/run_exp.sh.
  • Artifact-first verification and interpretation handoff: Checks that expected outputs exist under the session-scoped artifacts/ tree, then hands off to /interpret-experiment to generate a consolidated report.

Quick Start

Run /run-experiment after generating a session plan with /plan-experiment to execute the planned causal abstraction experiments and produce a session report.

Frequently Asked Questions about run-experiment

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I run causal abstraction experiments on language models end-to-end?▼

To run causal abstraction experiments, you execute a session plan by building validated task scaffolding, composing Hydra runner YAML configs, and running pipeline scripts to produce verified output artifacts without manual orchestration.

What is the process for orchestrating language model experiments using Hydra runner configs?▼

Experiment orchestration involves building and snapshotting Hydra runner YAML configurations in a session directory, running debug passes, and then executing the full preset pipeline via shell scripts to validate expected outputs.

How do I validate token alignment and forward-pass issues before executing a language model experiment?▼

You validate token alignment and forward-pass issues by optionally creating missing tasks and running task tests during the setup phase, catching these issues before the main pipeline execution begins.

Do I need a prior experiment plan to execute causal abstraction workflows?▼

Yes, you need a session plan generated beforehand. The execution workflow consumes research objective and plan documents from a session directory to run the specified experiment chain autonomously.

What happens after executing a causal abstraction experiment pipeline?▼

After execution, the system checks that expected outputs exist under a session-scoped artifacts tree, then delegates final writeout to an interpretation step to generate a consolidated report.

Can I run custom methods and analyses during mechanistic interpretability experiment execution?▼

Yes, the execution workflow supports including optional custom methods and analyses alongside sweep-driven runner runs when processing the causal abstraction experiment plan.

Related Skills

Explore More Education & Research Skills→
zivtechzivtech

study-design-planner

Operationalize research questions into PICO frameworks and causal DAGs.

Official
Advanced
zli5460zli5460
2

sparse-autoencoder-training

Train and analyze sparse autoencoders to interpret neural network activations.

Community
Intermediate
rethinkhealthrethinkhealth
6

ce:work-beta

Execute work plans end-to-end by parsing plan documents and orchestrating tasks, tests, and reviews.

Official
Advanced
EmaRimoldiEmaRimoldi

experiment-data-builder

Convert experiment-plan specifications into Hydra-backed dataset generators and loaders.

Community
Advanced