skills.rest
InverterNetwork
InverterNetworkOFFICIAL
Explore all skills by this author →
Website𝕏@inverternetworkGitHub

sparse-autoencoder-training

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

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

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 activate for multiple, unrelated concepts, making model behavior difficult to interpret.

Core Features & Use Cases

  • Feature Discovery: Decomposes dense model activations into sparse, monosemantic features using Sparse Autoencoders (SAEs).
  • Model Steering: Enables causal intervention by identifying and manipulating specific feature directions to influence model output.
  • Use Case: Researchers can use this to identify specific features responsible for harmful content or bias, then perform steering to mitigate these behaviors during inference.

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?▼

To decompose neural network activations into interpretable features, you train Sparse Autoencoders that resolve polysemanticity by converting dense activations into sparse, monosemantic feature directions. This skill manages the training configurations and reconstruction quality evaluation.

Can I use transformer-lens with sae-lens for mechanistic interpretability?▼

Yes, you can use transformer-lens with sae-lens for mechanistic interpretability. This integration allows you to process model activations, manage pre-trained SAE configurations, and identify top-activating features for specific input prompts.

How do sparse autoencoders help with model steering and mitigating bias?▼

Sparse autoencoders enable model steering by identifying specific feature directions responsible for harmful content or bias. You can then perform causal interventions during inference to manipulate these features and influence model output.

What is the best way to find specific features causing polysemanticity in neural networks?▼

The best way to find features causing polysemanticity in neural networks is using feature discovery workflows. This skill decomposes dense activations where individual neurons fire for unrelated concepts, isolating them into sparse, interpretable features.

Do I need pre-trained sparse autoencoders to analyze gpt2-small activations?▼

You do not need to train from scratch to analyze gpt2-small activations. You can load a pre-trained SAE for gpt2-small using sae-lens to immediately identify top-activating features for a given prompt.

Why does sparse autoencoder training require resolving polysemanticity?▼

Sparse autoencoder training requires resolving polysemanticity because individual neurons often activate for multiple unrelated concepts. Resolving this makes model behavior interpretable by mapping dense activations to sparse, monosemantic features.

Related Skills

Explore More Data & Analytics Skills→
RheasilviaRheasilvia

sparse-autoencoder-training

Train Sparse Autoencoders to decompose transformer activations into interpretable features.

Community
Advanced
cxnaivecxnaive

sparse-autoencoder-training

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

Community
Advanced
handsomelong922handsomelong922

sparse-autoencoder-training

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

Community
Intermediate
richardnguyen0715richardnguyen0715

sparse-autoencoder-training

Train and analyze Sparse Autoencoders for interpretable feature discovery in neural networks.

Community
Advanced