skills.rest
Plaidmustache
PlaidmustacheCOMMUNITY
Explore all skills by this author →
GitHub

sparse-autoencoder-training

Train and analyze sparse autoencoders on transformer activations with SAELens.

Updated May 4, 2026
Search Tags:#interpretability#feature-discovery#mechanistic-interpretability#sae#transformer-lens#sae-lens#sparse-autoencoders
One-click install
npx skills add https://github.com/Plaidmustache/hermes-nulab --skill sparse-autoencoder-training-plaidmustache
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: sparse-autoencoder-training
Source: https://github.com/Plaidmustache/hermes-nulab/tree/main/optional-skills/mlops/saelens
Command: npx skills add https://github.com/Plaidmustache/hermes-nulab --skill sparse-autoencoder-training-plaidmustache

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

This Skill provides practical guidance for training and analyzing Sparse Autoencoders (SAEs) with SAELens to decompose dense neural activations into sparse, interpretable features for mechanistic interpretability research.

Core Features & Use Cases

  • Training workflows to load pre-trained SAEs or train custom SAEs on transformer activations.
  • Analysis workflows to identify top-activating features per token, perform feature attribution and steering, and run ablation experiments.
  • Use cases include discovering interpretable features, studying superposition, and analyzing monosemantic representations in language models.

Quick Start

Load a pre-trained SAE for a GPT-2 small layer, encode activations for a sample prompt, and inspect the top activating features.

Frequently Asked Questions about sparse-autoencoder-training

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

FAQPage Schema
How do I train sparse autoencoders to find interpretable features in transformer models?▼

To train sparse autoencoders for transformer models, you can use SAELens to configure training runs on dense activations and identify monosemantic features. This workflow documents configuration, training, and evaluation steps for decomposing neural activations.

What is mechanistic interpretability and how do sparse autoencoders explain neural activations?▼

Mechanistic interpretability uses sparse autoencoders to decompose dense neural activations into interpretable, sparse features. This process helps identify monosemantic representations and study superposition in language models by mapping activations to distinct features.

How do I load a pre-trained SAE for GPT-2 and inspect top activating features?▼

Loading a pre-trained SAE for GPT-2 small involves using SAELens to fetch the model and encode activations for a sample prompt. You can then inspect the top activating features per token to analyze interpretable representations.

Can I perform feature steering and ablation experiments on transformer layers using SAELens?▼

Yes, you can perform feature steering and ablation experiments on transformer layers using SAELens. The workflow supports feature attribution, steering, ablation, and reconstruction quality checks across transformer models to analyze neural network behavior.

What are the limitations of using sparse autoencoders for feature discovery in language models?▼

Sparse autoencoders for feature discovery require careful configuration and evaluation of reconstruction quality. Limitations include ensuring features are truly monosemantic and validating that ablation or steering experiments accurately reflect the model's internal representations.

Related Skills

Explore More Education & Research Skills→
GregsGreyCodeGregsGreyCode
6

sparse-autoencoder-training

Train and analyze sparse autoencoders on transformer activations using SAELens.

Community
Advanced
MarbleSodasMarbleSodas

sparse-autoencoder-training

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

Community
Intermediate
Human-Frontier-Labs-IncHuman-Frontier-Labs-Inc

flow-nexus-neural

Train and deploy neural networks across distributed E2B sandboxes.

Official
Advanced
RheasilviaRheasilvia

sparse-autoencoder-training

Train Sparse Autoencoders to decompose transformer activations into interpretable features.

Community
Advanced