Training Sparse Autoencoders (opens in new tab)

Dictionary learning over transformer activations: what the trainer sees, how feature quality scales with data, and how a trained SAE plugs…