stdlogits/drowse-web-gemma3-4b-instruments

Gemma 3 4B instruments for drowse

The optional packs drowse.ai loads for stdlogits/drowse-web-gemma3-4b:

• R lens and S lens: drowse's own fits over the model's activations. The R lens reads what a vector disposes the model to say; the S lens, which input tokens it looks like it has read.
• Units: reads and layer contrasts for the model's neurons and attention heads, bound to those lenses and the core pack's neutral statistics.
• Gemma Scope features: Google DeepMind's Gemma Scope 2 sparse autoencoder for the residual stream after layer 17 (resid_post/layer_17_width_16k_l0_medium, 16k features), repackaged unchanged for the browser.

drowse fitted the lenses on FineWeb-Edu (sample-10BT at revision 87f09149); each lens and units pack's ATTRIBUTION.md credits the corpus, and its manifest records the exact corpus and its hash. The packs hold fitted matrices and statistics, not corpus text.

These are browser artifacts, not a Transformers checkpoint. Install them through drowse.ai: its signed catalog pins each download to an immutable revision of this repository and verifies every file's SHA-256 digest.

Licenses apply by component: see LICENSES.md.

drowse is written by a9lim with Claude; its browser edition, drowse.ai, is built from stdlogits/drowse.ai.