claudegoodies
Skill

sparse-autoencoder-training

From NousResearch

Provides guidance for training and analyzing Sparse Autoencoders (SAEs) using SAELens to decompose neural network activations into interpretable features. Use when discovering interpretable features, analyzing superposition, or studying monosemantic representations in language models.

Guides training and analyzing Sparse Autoencoders via SAELens to decompose model activations into interpretable features.

Use it when

  • Discovering interpretable features in language model activations
  • Studying superposition and polysemantic neurons
  • Training a custom SAE on a model's residual stream
  • Analyzing safety-relevant features like bias or deception

Skip it if

  • Need basic activation analysis without SAEs — use TransformerLens directly
  • Need causal intervention experiments — use pyvene or TransformerLens
  • Requires torch, transformer-lens, and sae-lens dependencies plus GPU for training

Facts

Repository
NousResearch/hermes-agent
Status
Actively maintained
Last commit
Source file
optional-skills/mlops/saelens/SKILL.md

Source preview

The instructions Claude Code reads when this skill runs.

# SAELens: Sparse Autoencoders for Mechanistic Interpretability

SAELens is the primary library for training and analyzing Sparse Autoencoders (SAEs) - a technique for decomposing polysemantic neural network activations into sparse, interpretable features. Based on Anthropic's groundbreaking research on monosemanticity.

**GitHub**: [jbloomAus/SAELens](https://github.com/jbloomAus/SAELens) (1,100+ stars)

## The Problem: Polysemanticity & Superposition

Individual neurons in neural networks are **polysemantic** - they activate in multiple, semantically distinct contexts. This happens because models use **superposition** to represent more features than they have neurons, making interpretability difficult.

**SAEs solve this** by decomposing dense activations into sparse, monosemantic features - typically only a small number of features activate for any given input, and each feature corresponds to an interpretable concept.

## When to Use SAELens

**Use SAELens when you need to:**
- Discover interpretable features in model activations
- Understand what concepts a model has learned
- Study superposition and feature geometry
- Perform feature-based steering or ablation
- Analyze safety-relevant features (deception, bias, harmful content)

**Consider alternatives when:**
- You need basic activation analysis → Use **TransformerLens** directly
- You want causal intervention experiment
View full source on GitHub →

Other skills