pulsatrix.feature_dashboard.v1: one learned feature (for example a sparse autoencoder unit) with its activation statistics and top-activating examples.
More...
#include <document.hpp>
pulsatrix.feature_dashboard.v1: one learned feature (for example a sparse autoencoder unit) with its activation statistics and top-activating examples.
- Note
- Follows the core of SAEDashboard and Neuronpedia's feature pages. Logit effects and other fields can be added later within v1.
◆ activation_density
| float pulsatrix::FeatureDashboardDocument::activation_density = 0.0f |
The fraction of inputs on which the feature is active, in [0, 1].
◆ feature_index
| int64_t pulsatrix::FeatureDashboardDocument::feature_index = 0 |
◆ histogram_counts
| std::vector<int64_t> pulsatrix::FeatureDashboardDocument::histogram_counts |
◆ histogram_edges
| std::vector<float> pulsatrix::FeatureDashboardDocument::histogram_edges |
Activation histogram: bin edges (one more than counts), or both empty.
◆ max_activation
| float pulsatrix::FeatureDashboardDocument::max_activation = 0.0f |
◆ source
| std::string pulsatrix::FeatureDashboardDocument::source |
Where the feature lives, for example "sae.blocks.3".
◆ top_examples
| std::vector<Example> pulsatrix::FeatureDashboardDocument::top_examples |
The documentation for this struct was generated from the following file: