You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Once #57 is completed, we can create UI in the Explore page to enable functionality when a dataset has SAE features associated with it.
There are few main buckets of functionality:
Filtering data points by features
Visualizing features of data points
Steering search via features
Filtering data points by features
We can provide an autocomplete feature selector that would then filter data points by the highest activations for the chosen feature. For an example of the feature selector see latent-taxonomy
We could also try transitioning the UMAP where the y-axis would then represent the activation of the data point for that feature.
Visualizing features of data points
We can show the top features activated for a data point in the table of results. For an example, see the visualizations in latent-taxonomy
We can also explore visualizations that surface the most commonly activated features for a selection of data points (e.g. for a cluster or a lasso selection)
Steering search via features
Currently, when we do search on our dataset we embed the text entered by the user and do a nearest neighbor search against the embeddings. One thing we could do is extract the features of the search text and then allow the user to turn them up or down. When the features have been modified we can modify the search vector before doing nearest neighbor search.
The text was updated successfully, but these errors were encountered:
Once #57 is completed, we can create UI in the Explore page to enable functionality when a dataset has SAE features associated with it.
There are few main buckets of functionality:
Filtering data points by features
We can provide an autocomplete feature selector that would then filter data points by the highest activations for the chosen feature. For an example of the feature selector see latent-taxonomy
We could also try transitioning the UMAP where the y-axis would then represent the activation of the data point for that feature.
Visualizing features of data points
We can show the top features activated for a data point in the table of results. For an example, see the visualizations in latent-taxonomy
We can also explore visualizations that surface the most commonly activated features for a selection of data points (e.g. for a cluster or a lasso selection)
Steering search via features
Currently, when we do search on our dataset we embed the text entered by the user and do a nearest neighbor search against the embeddings. One thing we could do is extract the features of the search text and then allow the user to turn them up or down. When the features have been modified we can modify the search vector before doing nearest neighbor search.
The text was updated successfully, but these errors were encountered: