sparse-autoencoder-training

Decompose dense neural network activations into sparse, interpretable features.

2|1|Updated Jul 14, 2026

Search Tags:#neural networks #superposition #feature discovery #sae #sparse autoencoders #mechanistic interpretability

Ready to run

npx skills add https://github.com/heysuhas/hermes_cli --skill sparse-autoencoder-training-heysuhas

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: sparse-autoencoder-training
Source: https://github.com/heysuhas/hermes_cli/tree/main/optional-skills/mlops/saelens
Command: npx skills add https://github.com/heysuhas/hermes_cli --skill sparse-autoencoder-training-heysuhas

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 difficult to interpret.

Core Features & Use Cases

Feature Discovery: Decompose dense model activations into sparse, monosemantic features that correspond to human-interpretable concepts.
Model Analysis: Study superposition and feature geometry to understand how models represent information.
Feature Steering: Identify and manipulate specific feature directions to influence model output or perform causal interventions.

Quick Start

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

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?▼

You decompose neural network activations into interpretable features by using sparse autoencoders to resolve polysemanticity, breaking dense activations into sparse, monosemantic concepts for mechanistic interpretability research.

How do I resolve polysemanticity in language models?▼

Resolving polysemanticity in language models involves decomposing dense overlapping neuron representations into sparse, monosemantic features using sparse autoencoders to isolate human-interpretable concepts.

Can I use transformer-lens with sparse autoencoders for feature discovery?▼

Yes, you can use transformer-lens with sparse autoencoders for feature discovery by integrating them with torch to analyze dense activations and reconstruct them into interpretable feature directions.

What is the best way to perform feature steering on model activations?▼

Performing feature steering on model activations requires decomposing dense representations into sparse features using sparse autoencoders, allowing you to identify and manipulate specific feature directions for causal interventions.

Do I need sae-lens to study superposition in language models?▼

Yes, studying superposition in language models requires sae-lens to load pre-trained sparse autoencoders, alongside transformer-lens and torch dependencies, to analyze feature geometry and dense activation behavior.

How do I load a pre-trained SAE for gpt2-small to identify activating features?▼

Loading a pre-trained SAE for gpt2-small uses sae-lens to identify top features activating on input text, requiring integration with transformer-lens and torch to perform activation analysis and reconstruction.