Verified:GitHub

sparse-autoencoder-training

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

1|Updated Feb 21, 2026

Search Tags:#feature discovery #sparse autoencoders #mechanistic interpretability #neural network analysis #transformer interpretability #sae lens

Ready to run

npx skills add https://github.com/tianhao909/AI-Research-SKILLs-cn --skill sparse-autoencoder-training-tianhao909

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: sparse-autoencoder-training
Source: https://github.com/tianhao909/AI-Research-SKILLs-cn/tree/main/04-mechanistic-interpretability/saelens
Command: npx skills add https://github.com/tianhao909/AI-Research-SKILLs-cn --skill sparse-autoencoder-training-tianhao909

SYSTEM DOCUMENTATION & REQUIREMENTS

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

What problem does it solve?

This Skill tackles the challenge of understanding complex neural network activations by decomposing them into sparse, interpretable features, making models more transparent.

Core Features & Use Cases

Feature Discovery: Identify and analyze interpretable concepts learned by neural networks.
Polysemanticity Analysis: Understand how models represent multiple concepts within single neurons.
Use Case: When debugging a language model's unexpected behavior, use this Skill to pinpoint the specific features in its activations that correspond to harmful content generation or bias.

Quick Start

Use the SAELens skill to load a pre-trained sparse autoencoder and encode model activations.

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 neural network activations?▼

This Skill provides tools and guidance to train sparse autoencoders, decomposing neural network activations into interpretable features to solve model transparency and polysemanticity issues.

What is polysemanticity and how do sparse autoencoders resolve it?▼

Polysemanticity is when a single neuron represents multiple concepts. Sparse autoencoders resolve it by decomposing activations into sparse, monosemantic features, making neural network concepts distinct and interpretable.

How can I analyze transformer activations for harmful content generation or bias?▼

You can analyze transformer activations for harmful content by using this Skill to pinpoint specific features corresponding to bias, enabling targeted feature analysis, attribution, and model steering.

Can I use SAELens and TransformerLens with PyTorch to steer language model behavior?▼

Yes, this Skill supports workflows for feature attribution and steering within language models, utilizing the SAELens and TransformerLens libraries alongside PyTorch to load autoencoders and encode activations.

What is the best way to load a pre-trained sparse autoencoder for model interpretability?▼

The best way to load a pre-trained sparse autoencoder is using the SAELens integration provided here, allowing you to quickly load models and encode activations for immediate feature discovery.