skills.rest
icyzh
icyzhCOMMUNITY
Explore all skills by this author →
Website𝕏@icyzhbeGitHub

sparse-autoencoder-training

Train sparse autoencoders to decompose transformer activations into interpretable features.

1|Updated Jul 31, 2026
Search Tags:#neural networks#feature discovery#sae#sparse autoencoders#mechanistic interpretability#transformer-lens
One-click install
npx skills add https://github.com/icyzh/hermes-web --skill sparse-autoencoder-training-icyzh
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: sparse-autoencoder-training
Source: https://github.com/icyzh/hermes-web/tree/main/optional-skills/mlops/saelens
Command: npx skills add https://github.com/icyzh/hermes-web --skill sparse-autoencoder-training-icyzh

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires sae-lens, transformer-lens, torch, and includes references (resource) components.

What problem does it solve?

This skill addresses the challenge of polysemanticity in neural networks, where individual neurons represent multiple, overlapping concepts, making model behavior opaque and difficult to interpret.

Core Features & Use Cases

  • Feature Discovery: Decompose dense model activations into sparse, monosemantic features that correspond to human-interpretable concepts.
  • Model Steering: Identify and manipulate specific feature directions to influence model output or perform causal interventions.
  • Use Case: Researchers can use this to identify features related to specific topics like code syntax or sentiment, then ablate or amplify those features to steer model behavior during generation.

Quick Start

Use the sparse-autoencoder-training skill to load a pre-trained SAE for gpt2-small and identify the top-activating features for a given input prompt.

Frequently Asked Questions about sparse-autoencoder-training

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

FAQPage Schema
How do I decompose neural network activations into interpretable features?▼

Sparse autoencoders resolve polysemanticity by decomposing dense neural network activations into sparse, monosemantic features that correspond to human-interpretable concepts. This technique isolates individual feature directions for clearer model analysis.

How do I identify and manipulate specific features for model steering?▼

Model steering uses sparse autoencoders to identify specific feature directions, then ablating or amplifying those features to causally intervene and influence transformer model output during generation without retraining.

Can I use transformer-lens and sae-lens to analyze high-dimensional activations?▼

Yes, transformer-lens and sae-lens support high-dimensional activation analysis by loading pre-trained SAEs for transformer models like gpt2-small to extract and reconstruct interpretable features from dense activations.

Why do individual neurons represent multiple overlapping concepts in neural networks?▼

Polysemanticity causes individual neurons to represent multiple overlapping concepts, making model behavior opaque. Sparse autoencoders address this by decomposing dense activations into sparse, monosemantic features for mechanistic interpretability.

Do I need torch to perform causal ablation experiments on transformer models?▼

Yes, torch is required alongside sae-lens and transformer-lens to perform causal ablation experiments on transformer models, providing the computational backend for high-dimensional activation analysis and feature manipulation.

Related Skills

Explore More Education & Research Skills→
InverterNetworkInverterNetwork

sparse-autoencoder-training

Trains and analyzes Sparse Autoencoders to decompose dense neural activations into sparse features.

Official
Advanced
cxnaivecxnaive

sparse-autoencoder-training

Train and analyze Sparse Autoencoders to decompose dense neural activations into sparse features.

Community
Advanced
brittaniebuffiecsubrittaniebuffiecsu
1

sparse-autoencoder-training

Train Sparse Autoencoders to decompose neural network activations into interpretable features.

Community
Advanced
GuitarAlchemistGuitarAlchemist

ix-supervised

Train and evaluate supervised machine learning models in Rust.

Official
Advanced