skills.rest
cesaregarza
cesaregarzaCOMMUNITY
Explore all skills by this author →
WebsiteGitHub

mechinterp-overview

Summarize SAE feature behavior with top tokens, activation statistics, and weapon associations.

1|Updated Jul 9, 2024
Search Tags:#overview#feature-analysis#pagerank#mechinterp#activation-stats#token-influence#sae
One-click install
npx skills add https://github.com/cesaregarza/SplatNLP --skill mechinterp-overview
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: mechinterp-overview
Source: https://github.com/cesaregarza/SplatNLP/tree/main/.claude/skills/mechinterp-overview
Command: npx skills add https://github.com/cesaregarza/SplatNLP --skill mechinterp-overview

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

This overview provides a fast, high-level summary of a SAE feature's behavior to help researchers decide which hypotheses to pursue. It highlights correlations (not causation) by presenting top tokens, activation statistics, family breakdown, weapon associations, and sample contexts, along with checks for existing labels and potential ReLU floor issues.

Core Features & Use Cases

  • PageRank-weighted top tokens indicating co-occurrence with high activation.
  • Activation statistics (mean, std, median, sparsity) and example counts to quantify feature behavior.
  • Family breakdown aggregating activations across ability families (SCU, SSU, etc.).
  • Top weapons associated with high-activation examples to reveal weapon-level patterns.
  • Sample high-activation contexts for quick inspection.
  • Existing label visibility and ReLU floor warnings to flag interpretability concerns.

Quick Start

To generate a quick overview, load the feature context and run the overview CLI or API for a given feature_id and model type, for example:

  • CLI: poetry run python -m splatnlp.mechinterp.cli.overview_cli --feature-id 18712 --model ultra
  • JSON: poetry run python -m splatnlp.mechinterp.cli.overview_cli --feature-id 18712 --model ultra --format json

Frequently Asked Questions about mechinterp-overview

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

FAQPage Schema
How do I get a high-level overview of an SAE feature to guide interpretability experiments?▼

Generate a fast SAE feature overview by running the CLI with a specific feature-id and model type, which outputs top tokens, activation statistics, and family breakdowns to guide your experiments.

What does PageRank-weighted token analysis reveal about SAE feature behavior?▼

PageRank-weighted top tokens indicate which tokens co-occur most frequently with high activations, helping you identify the specific contexts that trigger a feature.

Can I output SAE feature activation statistics and sample contexts in JSON format?▼

Yes, append the --format json flag to the overview CLI command to export activation statistics, sample high-activation contexts, and weapon associations as structured JSON.

Does this SAE overview tool identify ReLU floor issues and existing labels?▼

The overview includes visibility checks for existing labels and automatically flags potential ReLU floor issues to warn you of interpretability concerns during analysis.

When should I rely on SAE feature overviews for causal claims about model behavior?▼

Use these overviews for correlation insights and hypothesis generation only; they emphasize correlations rather than causal conclusions to guide exploratory experiments.

What statistics are included in an SAE feature overview for quantifying feature behavior?▼

The overview calculates mean, standard deviation, median, sparsity, and example counts to quantify feature behavior, alongside a family breakdown aggregating activations across ability families.

Related Skills

Explore More Data & Analytics Skills→
cesaregarzacesaregarza
1

mechinterp-crossmodel-matcher

Align SAE features between Ultra and Full models via top-token overlap.

Community
Intermediate
tianhao909tianhao909
1

transformer-lens-interpretability

Inspect and manipulate transformer internal activations and weights for mechanistic interpretability.

Community
Advanced
Exploration-labsExploration-labs
49

skill-performance-profiler

Analyze skill usage patterns to optimize token consumption across conversations.

Official
Intermediate
jaechang-hitsjaechang-hits
298

shap-model-explainability

Compute SHAP feature attributions and visualize model explainability.

Community
Advanced