sparse-autoencoder-training

Configure LanguageModelSAERunnerConfig with your model name, hook point, d_sae expansion factor, L1 coefficient, and training tokens, then run SAETrainingRunner. Enable l1_warm_up_steps and use_ghost_grads to prevent dead features during training.

Use SAE.from_pretrained with a release like gpt2-small-res-jb and a sae_id such as blocks.8.hook_resid_pre. Then encode activations from a TransformerLens HookedTransformer cache to get sparse feature activations.

Standard SAEs use ReLU with an L1 penalty for general-purpose sparsity. Gated SAEs add a learned gating mechanism for better sparsity control, while TopK SAEs activate exactly K features per input for consistent sparsity levels.

Dead features usually result from applying full L1 penalty too early. Set l1_warm_up_steps to 1000 or more, enable use_ghost_grads to revive inactive features, and avoid setting the L1 coefficient too high initially.

Yes, extract a feature direction from sae.W_dec and add it to the residual stream via a TransformerLens forward hook during generation. Scaling the direction by a strength parameter controls how strongly the feature influences output.

Use TransformerLens directly for basic activation analysis or causal intervention experiments that do not require sparse feature decomposition. SAELens is specifically for discovering interpretable features and studying superposition.