artifact
active
artifact:headvisHeadVis
Interactive tool for visualizing attention heads across the full data distribution; frontend open-sourced.
Neighborhood — ranked by edge-count
Papers (1)
paper
- HeadVis: An Interactive Tool For Investigating Attention Headsintroducesmentions
Methods (6)
method
- Feature Visualizationanalogous_toMethod of optimizing input to cause a neuron to fire maximally, used to characterize what a neuron detects; establishes causal link
- QK and OV attributionsimplementsWrites attention scores/head output as sums of feature interactions; core HeadVis capability.
- HeadVis PCA view of Q, K, O activationsimplementsHypothesis-generation tool for detecting distinct head behaviors via clustering.
- Custom-sequence inputimplementsServer feature letting users type sequences and view head attention; not open-sourced.
- Head ranking on a (query, key) pairimplementsRanks all heads by attention on a pair to trace composition.
- HeadVis head selector scatter plotimplementsConfigurable scatter plot of heads by metrics to find interesting heads.
Artifacts (2)
artifact
- HeadVis Gemma 3 1B demoimplementsHosted demo on all heads of Gemma 3 1B.
- HeadVis Haiku 3.5 demoimplementsHosted demo on a subset of Haiku 3.5 heads with cached attributions.
Datasets (1)
dataset
- Model whose heads are studied in the case studies.