Loading page
About this tool
Check label distribution in a classification dataset: imbalance ratio, entropy, balanced class weights and resampling counts.
The Dataset Class Balance Checker reads a list of classification labels and reports the imbalance ratio, normalised Shannon entropy, Gini impurity and the majority-class baseline accuracy for the dataset. It also prints scikit-learn's balanced class weights, n_samples / (n_classes x n_class_rows), and the row counts you would end up with after oversampling or undersampling. Labels are parsed in your browser from plain lines, a CSV column or a JSONL property, so no data leaves the page.
Open Dataset Class Balance Checker on AltFTool — it loads instantly in your browser.
Pick the 'Input format' - One label per line, CSV column or JSONL property - and paste your labels into the Labels box (CSV needs the label column plus the 'First row is a header' checkbox).
Set 'Cross-validation folds' (default 5) and the 'Case sensitive' toggle that decides whether 'Spam' and 'spam' are different classes.
Read the imbalance ratio with its severity label, majority-class baseline accuracy, entropy, Gini and per-class balanced weights, then press 'Copy report' for the tab-separated summary.
Parsing and maths happen client-side, so private datasets never leave the browser.
Copy the balanced class weights straight into class_weight or a loss function.
Warns about rare classes, unusable cross-validation splits and misleading accuracy.
An imbalance ratio at or below 1.5:1 behaves like a balanced dataset, and up to about 3:1 rarely needs special handling. Beyond roughly 10:1, plain accuracy stops being informative and you should switch to macro-F1, balanced accuracy or PR-AUC and consider resampling or class weights.
scikit-learn's class_weight="balanced" sets each class weight to n_samples / (n_classes x n_rows_in_class). With 100 rows split 90/10 across two classes, the majority weight is 100 / (2 x 90) = 0.56 and the minority weight is 100 / (2 x 10) = 5.0.
At minimum, each class needs as many rows as your cross-validation fold count, because StratifiedKFold cannot split a class with fewer members than n_splits (the default is 5). In practice fewer than about 10 rows in a class makes per-class precision and recall too noisy to compare between runs.
No. Resample only the training split and leave validation and test at the real-world distribution, otherwise your reported metrics describe a dataset that does not exist in production. Class weights are often preferred over oversampling because they leave the data untouched.