A quick read of your CSV before you submit, so you catch problems in seconds, not after a run of minutes to hours.
Your data stays in your browser - nothing is uploaded.
Drop a CSV here, or choose a file
Read locally in your browser. Nothing is uploaded. A quick sniff test, not a full run.
CSV
-
-
Auto-detected from your data. The protected attribute is always treated as categorical. Categorical columns are one-hot encoded (each distinct value adds one encoded column). Ignore columns are passed through unchanged and not debiased. Rosa cannot detect your outcome / label column - if your data has one (the value you would predict), tick it as Ignore yourself, or Rosa will modify it and break downstream model training. The protected attribute above is a guess from the column NAME, and on real datasets it is often the wrong column - a name containing "age" or "sex" wins even when the characteristic you care about is a race or income column named something else, and where no name matches at all the first categorical column is used. Check it and change it before you run: every proxy finding below is measured against whichever column is selected, so the wrong pick can report "no strong proxy found" on data that genuinely carries bias.
Column
Detected
Categorical
Ignore
What is driving the size
Encoded columns added by each categorical column
Categorical columns with many distinct values expand the most when one-hot encoded, and runtime grows with the encoded total. Numeric columns count as one each. Moving a high-cardinality or identifier column out of "categorical" (into ignore) drops its contribution from one-per-value to one. The total below is a conservative preflight encoded-width projection - it is computed the way the service estimates width before a run, and the exact width a real run records can differ.