sparse-autoencoder-training

To train a sparse autoencoder on transformer activations, configure a custom SAE using PyTorch-compatible activation hooks with sparsity controls, checkpointing, and evaluation metrics to decompose dense activations into interpretable features.

You can inspect which sparse autoencoder features activate for a specific prompt by loading a pre-trained SAE. This allows you to analyze and identify the most important sparse features representing concepts within your provided text tokens.

Yes, you can use SAELens and TransformerLens for feature steering and ablation. This workflow tests causal influence by amplifying or removing specific sparse features to observe how the transformer model's predictions change.

The best way to find features representing the same concept across different prompts is cross-prompt comparison. This analyzes sparse autoencoder activations to locate features consistently representing specific ideas like location, sentiment, or syntax.

Yes, you need PyTorch-compatible activation hooks to analyze pre-trained sparse autoencoders. These hooks encode, decode, reconstruct, and evaluate sparse features within transformer-based language models during inspection and attribution workflows.