Rebalance an imbalanced CSV label column

About this tool

Class Rebalancer evens out an imbalanced label column in a CSV by resampling whole rows — randomly duplicating rows from the smaller class(es) to bring them up (over-sampling) and/or randomly dropping rows from the larger class(es) to bring them down (under-sampling), toward a target class ratio you choose.

Pick the label column and a strategy, and every class moves toward the same size (or a ratio you set). The picks are driven by a fixed seed, so the same input and seed always produce the same output — change the seed for a different draw. Everything runs locally in your browser; no rows are uploaded, and no synthetic rows are invented (this is random resampling, not SMOTE).

Worked example

CSV — one spam row and three ham rows (a 1:3 imbalance):

text,label
buy now,spam
hello,ham
hi there,ham
see you,ham

With label column label, strategy oversample, and target ratio 1.0, the minority spam class is duplicated up to the majority size (3), keeping the originals first and appending the duplicates:

text,label
buy now,spam
hello,ham
hi there,ham
see you,ham
buy now,spam
buy now,spam

Switch Output to the before/after summary to get a JSON count report instead of the CSV:

{
  "classes": [
    {"label": "spam", "before": 1, "after": 3},
    {"label": "ham", "before": 3, "after": 3}
  ],
  "strategy": "oversample",
  "target_ratio": 1,
  "total_before": 4,
  "total_after": 6
}

Limits & edge cases

• Oversample only grows classes (duplicates minority rows); undersample only shrinks classes (drops majority rows); combine moves every class to a common size, growing the small ones and shrinking the big ones. Auto is the same as oversample.
• Target ratio is the desired minority-to-majority ratio after resampling, from just above 0 to 1.0. 1.0 = fully balanced (every class equal); 0.5 = the smaller class ends at half the larger. Values of 0 or above 1 are rejected.
• The label column can be a header name (when the header box is on) or a 1-based column number. Blank means the last column.
• You need at least two distinct classes in the label column — a single-class column has nothing to rebalance.
• Over-sampling duplicates existing rows verbatim, so it cannot add new information; it just reweights the classes. A very small target ratio combined with over-sampling can request an unrealistically large table, which is capped.
• With the same seed, the output is identical every time. Turn on shuffle to interleave the duplicates instead of appending them at the end.

FAQ