Encodes the prompt's last-token residual at the SAE's layer and ranks the firing features. TopK SAE has at most K=50 nonzero features per token, so request up to ~50.

Runs model.generate twice: once with no hooks, once with all active sliders applied as residual-stream additions h ← h + α · W_dec[:, feat].

Each point is one SAE feature; positions are the top-3 PCA components of W_enc. Click a point to add it as a steering target.

Encodes each prompt's last-token residual through the SAE. Returns per-sample top features and corpus-level firing rates.

Encodes both sets and ranks features by |fire_rate(A) − fire_rate(B)|. Shows which features distinguish A from B.

Encode docs, then filter by feature signature.

"Includes" = keep only docs where this feature fired (any activation). "Excludes" = drop docs where it fired.

For each prompt in the corpus above, generate a steered completion with h ← h + α · W_dec[:, feature]. Useful for producing targeted training examples that fire a specific feature.