Hand in any movie review. A DistilBERT model reads it in your browser, stamps a verdict on it, and marks up the sentences that decided it. Never trained on IMDB, it gets – of 25,000 IMDB reviews right on its own, and – with a small word-counting model trained on IMDB beside it.
or clip one of these:
The model downloads once (about 67 MB) the first time you stamp a review.
Your review comes back here, marked up.
The scoreboard
Accuracy on IMDB's 25,000 test reviews, half positive and half negative. Lines show the 95% interval. Every choice was made on 2,000 held-out training reviews first; the test set was scored once.
Long reviews
Hollow red rings are reviews over the model's 512-token window; solid dots are the rest.
short reviewslong reviews
two different readers
is the gap real?
where it still fails
Sarcasm. "Truly the cinematic event of the decade" reads as praise to a model that has learned praise words. Try the sarcasm clipping above.