Loading…
Loading…
Open a neuron and press Add to mix, then tune its dose. The recipe uses the neuronary format.
Generation runs in app.py / the molab notebook — this static page has no model.
Here is where a neuron fires hardest (the boxed token is its peak). Which label did the labeler give it?
DuckDB-WASM over the current model's parquet files. Tables: neurons, labels, steer, tokens, examples (all layers).
Each point is one MLP neuron, placed by its output direction (UMAP of the down-projection column), so neurons that write similar things sit together. Fires on: the corpus documents with the highest activation, stained by activation per token. Promotes: the logit lens of the neuron's output direction. Pushes: replies with the neuron's activation raised or lowered at every token, by a dose in units of the median residual-stream norm. Labels and the held-out check come from an LLM labeler, so treat them as hypotheses.