You bring three files
Train and test must have the same columns in the same order. The unlabelled file is the same
again, minus the target column.
Splitting is yours to do. Shuffle first, then cut. Do not split on something ordered (a date, an
ID range), or your test set will not represent your training set.
The test file keeps its labels. That is not a mistake in the docs. The platform needs the answers
to mark itself; it never learns from them.
Column layout
The unlabelled file is identical without the last column:
The target column
Binary or multi-class. Values can be numbers or text;0/1, N/Y, good/bad and
fraud/legit all work. Use the same values in train and test.
Feature Refinery handles binary targets only today. A multi-class dataset goes straight to
Model Crucible.
Sizes and limits
Feature Refinery’s column limit is not arbitrary: every column is a full round of training, so
width is what sets the run length.
Mistakes that cost the most time
Leaving the labels out of the test file. The run cannot score itself and will refuse the input. Columns in a different order between the two files. Matching is positional, not by name. Same order, both files. An ID that is not unique. Use a real key, or generate one. Row numbers are fine as long as they do not repeat. A column that leaks the answer. Adecision_date that only exists on approved applications, or
a case_closed_reason written after the outcome was known, will give you a model that looks
excellent and predicts nothing useful in production. Ask of every column: would I have had this
value at the moment I needed the prediction?
Ordered rows. Shuffle before splitting.
A near-constant column. A column that is the same value in 99% of rows adds nothing. Feature
Refinery will tell you, but you can save it the round.
When you are not sure which columns to include
Include them. That is what Feature Refinery is for. Bring a wider set than you are confident about and let the run tell you which ones earned their place, up to the 70-column limit. Guessing in advance is how useful signal gets thrown away.Sample files
The German Credit dataset, used throughout these docs:- german-credit-train.csv — 800 rows, 21 features, target
class - german-credit-test.csv — 200 rows, labels kept
- german-credit-unlabelled.csv — 200 rows, no target
See them used end to end
The German Credit walkthrough.