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

mechinterp-crossmodel-matcher

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

1|Updated Jul 9, 2024
Search Tags:#cross-model#interpretability#activation-patterns#SAE#feature-matching#token-overlap#model-scale
One-click install
npx skills add https://github.com/cesaregarza/SplatNLP --skill mechinterp-crossmodel-matcher
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: mechinterp-crossmodel-matcher
Source: https://github.com/cesaregarza/SplatNLP/tree/main/.claude/skills/mechinterp-crossmodel-matcher
Command: npx skills add https://github.com/cesaregarza/SplatNLP --skill mechinterp-crossmodel-matcher

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

This skill helps researchers align and compare SAE features across Ultra and Full models, revealing corresponding patterns and discrepancies to support interpretability.

Core Features & Use Cases

  • Cross-model feature mapping: Identify which Ultra features map to Full features based on activation and token patterns.
  • Similarity-aware validation: Quantify overlaps and detect monosemantic representations across model scales.
  • Use Case: Validate that a pattern discovered in Ultra has a counterpart in Full to ensure robust interpretation across model sizes.

Quick Start

To start, load Ultra and Full SAE contexts and run the FeatureMatcher to locate cross-model matches.

Frequently Asked Questions about mechinterp-crossmodel-matcher

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

FAQPage Schema
How do I align SAE features between different model scales?▼

Cross-model interpretability reveals corresponding patterns and discrepancies across model scales by comparing SAE feature activations and token patterns to validate robust interpretation.

How do I validate SAE feature mappings between an Ultra and a Full model?▼

You can validate cross-model feature mappings by quantifying top-token overlaps between Ultra and Full SAE contexts to detect corresponding monosemantic representations.

Can I check SAE feature monosemanticity across different model sizes?▼

Yes, you can check monosemanticity across model sizes by computing top-token overlap between Ultra and Full SAE contexts to verify if discovered patterns have cross-scale counterparts.

Do I need pre-loaded SAE contexts to perform cross-model feature matching?▼

Yes, you must load both Ultra and Full SAE contexts before computing top-token overlaps and generating the ranked cross-model match report with matching notes.

What is the best way to find cross-model SAE feature correspondences?▼

The best way to find cross-model SAE correspondences is to automatically align features by computing top-token overlap between loaded Ultra and Full SAE contexts to produce a ranked match report.

Related Skills

Explore More Data & Analytics Skills→
cesaregarzacesaregarza
1

mechinterp-cluster-mapper

Analyze cross-feature activations to identify SAE subsystems and co-activation patterns.

Community
Intermediate
MarbleSodasMarbleSodas

sparse-autoencoder-training

Train and analyze Sparse Autoencoders with SAELens to decompose neural activations into interpretable features.

Community
Intermediate
cloudlinesscloudliness

sparse-autoencoder-training

Train and analyze Sparse Autoencoders with SAELens on transformer activations.

Community
Advanced
OpenLAIROpenLAIR
1.0k

transformer-lens-interpretability

Analyze and steer transformer activations with TransformerLens for mechanistic interpretability.

Official
Advanced