I have a tabular dataset from UCI Adult Census Income with the following columns:

  - age (number)
  - workclass (category)
  - education (category, ordered from preschool through doctorate)
  - education-num (number, 1-16)
  - marital-status (category)
  - occupation (category, 14 unique values)
  - relationship (category)
  - race (category)
  - sex (binary: Male / Female)
  - capital-gain (number, heavily skewed, mostly zero)
  - capital-loss (number, similar to capital-gain)
  - hours-per-week (number, 1-99)
  - native-country (category, high cardinality ~40 classes)

The target column is "income" (binary: >50K or <=50K).

The dataset has about 48k rows. Training should be reasonably fast — prefer the
medium_quality preset. Use the concat combiner with two FC layers. Use AdamW with a
learning-rate scheduler.
