sparse-autoencoder-training

1# SAELens: Sparse Autoencoders for Mechanistic Interpretability
2 
3SAELens 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.
4 
5**GitHub**: [jbloomAus/SAELens](https://github.com/jbloomAus/SAELens) (1,100+ stars)
6 
7## The Problem: Polysemanticity & Superposition
8 
9Individual 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.
10 
11**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.
12 
13## When to Use SAELens
14 
15**Use SAELens when you need to:**
16- Discover interpretable features in model activations
17- Understand what concepts a model has learned
18- Study superposition and feature geometry
19- Perform feature-based steering or ablation
20- Analyze safety-relevant features (deception, bias, harmful content)
21 
22**Consider alternatives when:**
23- You need basic activation analysis → Use **TransformerLens** directly
24- You want causal intervention experiments → Use **pyvene** or **TransformerLens**
25…