The summary.
ALWAYS USE A PIPELINE OBJECT FOR PREPROCESSING
It is what prevents leakage, since transformers are fitted only on training folds. Fitting a scaler outside one leaks test information.
And use the time-series split for ordered data, never the default.
A NOTEBOOK CAN WORK WHILE ITS CODE, RUN FRESH, DOES NOT
Hidden state from deleted cells is why. Restart and run everything periodically, move reusable logic into modules, and never schedule a notebook as a production job.
PIN EXACT VERSIONS AND USE A CONTAINER FOR ACCELERATED HARDWARE
Driver and library compatibility is where most failures occur, and version ranges resolve differently over time.
CHECK THE LICENCE BEFORE USING A PUBLISHED MODEL
Some permit research only. And loading a model file can execute arbitrary code — prefer formats that cannot, and publishers you can identify.
Pin the model version; published models are updated and behaviour changes.
LOW DEVICE UTILISATION USUALLY MEANS THE DATA PIPELINE, NOT THE MODEL
And optimiser state can exceed the model's own size several times over, which is what memory estimates usually miss.
NEVER OVERWRITE A MODEL ARTEFACT
Any prediction attributed to it becomes unexplainable.
