sparse-autoencoder-training

Train sparse autoencoders using SAELens to decompose transformer activations into interpretable features. This process evaluates sparsity and reconstruction quality, enabling mechanistic insights and safer model deployment for GPT-2 style transformers.

Sparse autoencoder feature discovery identifies interpretable representations from transformer activations using SAELens. By training custom SAEs, it decomposes hidden states into sparse, causally relevant features for mechanistic interpretability.

Yes, analyzing GPT-2 activation features requires Python 3.10+, TransformerLens, and SAE tooling. This environment setup allows you to load pre-trained SAEs and inspect activation features across GPT-2 style models.

Yes, you can steer transformer model outputs by analyzing, steering, and ablating sparse autoencoder features. This technique helps study causal contributions of specific interpretable features extracted from the model activations.

Evaluate sparsity and reconstruction quality during custom SAE training by analyzing how effectively the sparse autoencoders decompose transformer activations. SAELens provides configurations to assess feature quality and reconstruction fidelity.

Load pre-trained sparse autoencoders using SAELens to immediately inspect activation features in GPT-2 style models. This allows you to bypass custom training and begin analyzing existing interpretable feature representations directly.