An autoresearch loop proposes Jev questions about wine reviews, converts the answers to numeric features, and fits a CatBoost regressor against human scores. Subsequent rounds use model errors to revise the features.
Original by TypeSafe AIEvaluationTypeSafe cookbookSource reviewed
Before you dive in
What you’ll find in the original
Keep the prediction target and held-out evaluation separate from feature proposals.
Turn Choice and Score answers into columns a conventional supervised model can use.
Compare the resulting regressor with a direct Jev score rather than assuming more questions improve it.
Worth knowing
Its measured improvement belongs to this wine-review dataset; the loop does not replace labeled data.