Skip to main content

Notes

Short write-ups of things I learned building the projects, with the numbers and the code behind them.

  1. · ML, Evaluation, Data

    Oversampling before the split made my model look better than it was

    On a cropland project, SMOTE ran before the train-test split. The score dropped once I measured on untouched points, and the baseline put it in context.

  2. · ML, MLOps, Evaluation

    Scoring a LightGBM model without the LightGBM library

    The fraud demo's first deploy failed on a missing system library. Reading the model file directly fixed it, and tests show the scores didn't change.