Out-of-fold (honest, cross-fit) predictions on the 773 train wells, eval-zone rows only (where
TVT_input is NaN — the rows that get scored). CV = GroupKFold(5) on WELLNAME (no well leakage): each
well's prediction comes from a model trained on the OTHER 4 folds.
Engine: per-well affine typewell-GR recalibration (GR_well ≈ g·GR_tw(TVT)+b, Huber IRLS on the known
zone, shrunk by λ) applied inside the particle… See the full description on the dataset page:
https://huggingface.co/datasets/replikeit/rogii-oof-claude.
