Gemma 3 1B / recorded responses
What does this unit respond to?
Choose a unit. Try different words. A preference is a clue, not a definition.
Loading recorded activations…
Each dot is one word. Click a category to inspect its words. The larger mark is the category mean. Axis fits each unit.
Inspect animal words
What am I looking at?
A “neuron” here is one coordinate of the gated MLP vector before down_proj. Layers run from 0 to 25. Unit IDs are local to each layer; moving deeper selects a different unit.
All 320 words were measured in They mentioned the word {TARGET}. There are 40 words in each category. Standardized responses use this unit’s mean and population SD across all 320 words. Zero means its average response, not inactivity. Larger values do not mean a unit is more important.
The 20 frozen winners were selected on a separate animal/tool vocabulary. Extra units let you explore all layers: the three strongest animal/tool contrasts and the contrast closest to zero in each layer. These extras are exploratory, not a new held-out claim. A near-zero animal/tool contrast does not imply an inactive unit.
This is a browser for saved measurements. Controls select units and words; they do not alter activations, silence neurons, or run the model. Download the data and provenance.