use case · wild
How few words does it take to swing the verdict? This tool scores every word's pull with occlusion attribution, then greedily removes the words the model leans on — the biggest pullers toward the current label — until the sentiment flips. You get the minimal "load-bearing" words laid bare, and a rewritten sentence that crosses the line.
Starts as
Word pulls toward the starting verdict (occlusion attribution):
This is a search over the model's own predictions, not a language rewrite — it removes words rather than replacing them, so the result reads like a redaction. It's a vivid way to see which tokens carry the sentiment: strip "delicious" and "wonderful" and a rave can collapse to neutral or negative.