mechinterp-patterns

Identify recurring SAE feature patterns and apply standardized labels.

1|Updated Jul 9, 2024
One-click install
npx skills add https://github.com/cesaregarza/SplatNLP --skill mechinterp-patterns
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: mechinterp-patterns
Source: https://github.com/cesaregarza/SplatNLP/tree/main/.claude/skills/mechinterp-patterns
Command: npx skills add https://github.com/cesaregarza/SplatNLP --skill mechinterp-patterns

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

This Skill provides a centralized reference of recurring SAE feature patterns to guide consistent labeling decisions during MechInterp investigations and feature labeling tasks.

Core Features & Use Cases

  • Pattern Catalog: A growing repository of recurring pattern archetypes with clear signatures and recommended labels.
  • Quick Pattern Reference: A structured index to map new features to established patterns, ensuring terminology consistency.
  • Update Guidance: Step-by-step workflow to add, review, and promote new patterns across the team.
  • Use Case: When encountering a new SAE feature, consult the catalog to match a pattern and apply the standard label.

Quick Start

Open the main patterns document at /mnt/e/mechinterp_runs/potential_patterns.md and search for a matching pattern. If found, label the feature using the catalog's standardized name. If not, follow the workflow to propose a new pattern.

Frequently Asked Questions about mechinterp-patterns

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

FAQPage Schema
How do I maintain consistency when labeling SAE features during mechinterp investigations?

To maintain consistency when labeling SAE features, consult a centralized catalog of recurring pattern archetypes to match new features and apply standardized terminology. This ensures team-wide alignment during mechinterp investigations.

What is the workflow for proposing a new SAE feature pattern?

Proposing a new SAE feature pattern involves following a structured update workflow to add, review, and promote new pattern entries within the central patterns document. This maintains a growing repository of validated feature archetypes for team use.

How do I map a new SAE feature to an existing pattern catalog?

To map a new SAE feature, open the main patterns document and search the structured index for matching signatures. If a matching pattern is found, apply the catalog's standardized name to label the feature consistently.

Do I need a specific file path to start cataloging SAE patterns?

Yes, you need access to the central patterns document located at /mnt/e/mechinterp_runs/potential_patterns.md. This file serves as the required repository for querying existing patterns and updating new feature archetypes.

What should I do if I cannot find a matching pattern for an SAE feature?

If you cannot find a matching pattern for an SAE feature, you should follow the defined update workflow to propose a new pattern. This involves adding the new feature signature to the catalog for review and standardized labeling.