Jev turns 2,000 wine notes into CatBoost features at 1.77 RMSE

Result

Held-out RMSE 1.77 after five rounds, versus 3.09 for the mean baseline

Ask Jev Score and Noul questions that turn tasting notes into numeric features for CatBoost.

This Jev research use case starts from 2,000 wine tasting notes and a critic score. CatBoost can only train on numbers, so the loop has to invent Jev questions that turn each note into a row of answers.

Each round Claude writes Score and Noul questions. Jev answers them for every note. CatBoost guesses the critic score. The 30 worst misses and 30 best hits go back so Claude can add, reword, or drop questions.

Round 1 used 18 questions and reached 1.87 RMSE on 800 held-out reviews. After five rounds the set is 38 questions and the held-out RMSE is 1.77, against 3.09 for guessing the mean. The mill on the left is that loop. The original source is the cookbook.