sparse-autoencoder-training

To train a sparse autoencoder for interpretable feature discovery, you use this Skill to guide the training process on neural network activations, decomposing them into monosemantic representations. It requires Python 3.10+, sae-lens, transformer-lens, and torch.

Superposition analysis studies how neural networks represent more features than they have dimensions by analyzing feature geometry. This Skill trains sparse autoencoders to decompose these superposed activations into distinct, interpretable features for language models.

Yes, analyzing monosemantic representations requires transformer-lens and sae-lens to extract and process model activations. You also need PyTorch version 2.0.0 or higher and Python 3.10 or above to run the training and analysis scripts.

Yes, you can identify concepts a language model has learned by training a sparse autoencoder on its activations. This process decomposes the internal representations into interpretable features, revealing the specific concepts the model uses for reasoning.

The best way to study feature geometry is to train sparse autoencoders on model activations to separate superimposed features. This Skill provides scripts and assets to analyze the resulting monosemantic representations and geometric properties.

Your sparse autoencoder training scripts may fail if your environment lacks the required dependencies or uses an incompatible Python version. Ensure you have Python 3.10+ installed alongside sae-lens>=6.0.0, transformer-lens>=2.0.0, and torch>=2.0.0.