How to use it
- Click a word in the analyzed text below. You see the features active on it and a graph of why the model wrote that word.
- Click a feature to see where else it fires, in thousands of real stories.
- Type your own text and press Analyze, or search for a feature by word.
Click a token to see the features active on it.
Features most active in this text
Find a feature
Each feature is removed in turn (its contribution subtracted from the internal state at its position) and the model is rerun: a blue edge means the token's score drops without the feature (it supported the token), red means it rises. Edge width is the size of that change. Grey edges show which token each feature was active on. Click a feature to inspect it.
Token itself
Previous token
Next token
Strongest activations on validation stories
How to read this: the SAE was trained on the residual stream entering block ?
of the model. Each feature has a label made automatically from its strongest examples: "ball"
means it fires on that token, after "named" on the token after it, before "said" on
the token before it. The percentage is how many of those examples agree, so a low percentage means the
label does not explain the feature well. Labels and examples come from a script
(python -m glassbox.features), not from a person.
The first position (the start-of-story token) is skipped, as in training. Shading shows activation relative to the feature's strongest validation example.