Tangerine Prep Research

Chev 1.0

A knight-errant for your weak spots.

Chev is a causal transformer that reads a learner as one time-ordered stream of question attempts and predicts, for any question they have not seen, how likely they are to get it wrong. It powers personalized weak-spot recommendations across Tangerine Prep.

How it scores today

0.789
Per-user AUC
honest accuracy on learners never seen in training
0.803
Pooled AUC
ranking quality across all held-out attempts
0.505
Log loss
calibration of the probabilities
894
Learners
in the training cohort

Playground

Rank a learner's hardest unseen questions. Requires an API key.

Chev 1.0 · served from chev.tangerineprep.app · /api/stats