sparse-autoencoder-training

You can decompose neural network activations into interpretable features by training Sparse Autoencoders (SAEs). This Skill provides workflows to load pre-trained SAEs or train custom ones to resolve polysemanticity in language models.

Yes, you can analyze pre-trained SAEs for GPT-2 small. The Skill supports using transformer-lens and sae-lens to load pre-trained SAEs and perform feature attribution on the model's internal representations.

To train a custom sparse autoencoder for mechanistic interpretability, this Skill provides scripts and guidance for configuring hyperparameters. It facilitates feature discovery and helps understand learned concepts in large language models.

The best way to identify features responsible for biased text generation is by decomposing the model's internal representations using Sparse Autoencoders. This approach pinpoints the underlying features driving specific behaviors in language models.

Yes, this skill requires PyTorch, sae-lens, and transformer-lens dependencies. These frameworks provide the foundational environment needed to train custom SAEs and perform feature steering on large language models.

Language models exhibit polysemanticity in dense activations because individual neurons represent multiple overlapping concepts. Using Sparse Autoencoders resolves this by decomposing the dense activations into sparse, interpretable features.