skills.rest
Signmanal
SignmanalCOMMUNITY
Explore all skills by this author →
GitHub

sparse-autoencoder-training

Decompose polysemantic transformer activations into sparse interpretable features with SAELens.

1|Updated Jun 25, 2026
Search Tags:#transformer models#superposition#feature discovery#sparse autoencoders#mechanistic interpretability#sae-lens#activation analysis
One-click install
npx skills add https://github.com/Signmanal/VIGIL --skill sparse-autoencoder-training-signmanal
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: sparse-autoencoder-training
Source: https://github.com/Signmanal/VIGIL/tree/main/optional-skills/mlops/saelens
Command: npx skills add https://github.com/Signmanal/VIGIL --skill sparse-autoencoder-training-signmanal

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

Neural network activations are polysemantic due to superposition, making it difficult to interpret what concepts language models have learned. This Skill uses Sparse Autoencoders (SAEs) to decompose dense activations into sparse, monosemantic features that correspond to human-interpretable concepts.

Core Features & Use Cases

  • Pre-trained SAE Analysis: Load and analyze existing SAEs to identify which features activate for specific inputs.
  • Custom SAE Training: Train your own SAEs on target model layers with configurable sparsity and architecture options.
  • Feature Steering & Ablation: Modify model behavior by amplifying or removing specific interpretable features for causal analysis.
  • Use Case: AI safety researchers can use this Skill to discover features related to deceptive behavior or bias in language models for auditing purposes.

Quick Start

Use the sparse-autoencoder-training skill to identify the top 5 activating interpretable features for the input "The Eiffel Tower is located in Paris" using a pre-trained GPT-2 small SAE.

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 polysemantic neural network activations into interpretable features?▼

To decompose polysemantic neural network activations, you use Sparse Autoencoders to project dense activations into sparse, monosemantic features that correspond to human-interpretable concepts, solving the superposition problem in transformer models.

How do I train a custom sparse autoencoder on transformer model activations?▼

You can train a custom sparse autoencoder on target transformer model layers by configuring sparsity and architecture options using SAELens and TransformerLens in PyTorch, allowing you to load pre-trained checkpoints or build new SAEs.

What is mechanistic interpretability and how does feature discovery work for language models?▼

Mechanistic interpretability uses feature discovery to identify which sparse autoencoder features activate for specific inputs, revealing human-interpretable concepts within polysemantic language model activations for safety-aligned model auditing.

Can I use sparse autoencoders for feature steering and ablation in AI safety research?▼

Yes, sparse autoencoders support feature steering and ablation by amplifying or removing specific interpretable features, enabling causal analysis to discover deceptive behavior or bias in transformer language models for auditing purposes.

Does this sparse autoencoder workflow require SAELens and TransformerLens?▼

Yes, this workflow requires SAELens, TransformerLens, and PyTorch to train custom SAEs, load pre-trained checkpoints, and perform feature steering or ablation experiments on transformer language model activations.

Why are transformer model activations polysemantic and difficult to interpret?▼

Transformer model activations are polysemantic due to superposition, where multiple features overlap in dense activation vectors, making it difficult to interpret learned concepts without decomposing them into sparse, monosemantic features.

Related Skills

Explore More Education & Research Skills→
dawsonblockdawsonblock

sparse-autoencoder-training

Train sparse autoencoders on transformer activations with SAELens.

Community
Advanced
cxnaivecxnaive

sparse-autoencoder-training

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

Community
Advanced
GregsGreyCodeGregsGreyCode
6

sparse-autoencoder-training

Train and analyze sparse autoencoders on transformer activations using SAELens.

Community
Advanced
OpenLAIROpenLAIR
1.0k

sparse-autoencoder-training

Train sparse autoencoders to decompose transformer activations into interpretable features.

Official
Advanced