// split rows into train / val / test by ratio
Split your dataset rows into train, validation, and test sets instantly. Enter total rows and ratio, get exact counts for ML workflows. Free, browser-based.
Ready to split
Enter total rows and ratio, then click CalculateType the total number of samples in your dataset (e.g. 10,000).
Enter train/val/test ratios or click a preset. Ratios are auto-normalized.
View exact row counts, percentages, and copy the generated Python code.
A dataset split calculator divides your total row count into separate subsets — train, validation, and test — based on a ratio you specify. This is a critical step in any ML pipeline to prevent data leakage and measure real model generalization.
The most common split is 70/15/15 or 80/10/10. For large datasets (1M+ rows), even 90/5/5 is sufficient since 50,000 rows for validation is still plenty. Smaller datasets may benefit from 60/20/20 to ensure robust evaluation.
Not always. If you're using cross-validation (k-fold), you can skip a dedicated validation set and just split into train/test (e.g. 80/20). A separate validation set is most useful when tuning hyperparameters or early-stopping during training.
This tool uses rounding, which can cause ±1 row discrepancy. The train set absorbs any remainder to ensure the total is always exact. This is standard practice in frameworks like scikit-learn's train_test_split.
If your ratios don't sum to 100 (e.g. 7/1.5/1.5 or 2/1/1), the tool automatically normalizes them to proportions. So 7/2/1 becomes 70%/20%/10%. This lets you enter natural ratios without needing to calculate percentages first.
Yes — almost always. Random shuffling prevents bias if your data is ordered (e.g. time-based, by class label). The generated Python code uses sklearn.model_selection.train_test_split with shuffle=True by default.
Yes. Set the validation ratio to 0 and split between train and test only. The validation card will show 0 rows and be visually dimmed. The generated Python code will reflect a simple two-way split.
A dataset split calculator helps data scientists and machine learning engineers determine exactly how many rows to allocate to each subset of a dataset — typically train, validation, and test. Rather than eyeballing percentages and doing mental math, this tool instantly computes exact integer row counts based on your chosen ratio, handles rounding automatically, and outputs code you can paste directly into your Python workflow.
Whether you're working with a 1,000-row CSV or a 50-million-row database export, the math is the same: divide by ratio, round carefully, and verify the total adds up. This tool does all of that in milliseconds, with a visual breakdown to boot.
💡 Looking for premium web development assets? MonsterONE offers unlimited downloads of templates, UI kits, and developer assets — worth checking out.
The foundation of any trustworthy ML model is a properly partitioned dataset. Without a held-out test set, you can't evaluate whether your model generalizes to unseen data — you're just measuring how well it memorized training examples. The three-way split (train/val/test) is now standard in the field:
The right split ratio depends heavily on your dataset size, model complexity, and evaluation requirements:
The calculation follows a specific order to ensure the total never exceeds or falls short of your dataset size. Here's the exact algorithm:
test = total − train − valThis mirrors the behavior of sklearn.model_selection.train_test_split and similar utilities, so the numbers you get here will match what you see in practice.
After calculating your split, this tool generates ready-to-use Python code using scikit-learn. You can paste it directly into a Jupyter notebook or .py script. The code uses train_test_split twice — first to carve out the test set, then again to split the remainder into train and validation. This is the canonical approach and avoids data leakage.
If you set validation to 0, the code outputs a simpler single train_test_split call, which is appropriate when using cross-validation strategies like KFold or StratifiedKFold for hyperparameter tuning.
For classification tasks, a simple random shuffle can result in class imbalance across splits — especially with rare classes. In those cases, use stratify=y in scikit-learn to ensure each split has the same class distribution as the full dataset. The row counts calculated here remain valid regardless of stratification; only the row selection logic changes.
For regression or unsupervised tasks, a simple random shuffle is usually sufficient. The important thing is that the split is reproducible — always set a random_state seed.
Standard random splitting is not appropriate for time-series datasets. Randomly shuffling temporal data causes data leakage because future data points end up in the training set. For time series, splits should be chronological: the first N% of time steps go to training, the next M% to validation, and the final K% to testing. The row counts this tool calculates are still useful — just apply them sequentially rather than randomly.
A practical rule of thumb: your test set should have at least 1,000 samples to produce statistically reliable evaluation metrics. For a binary classifier, you want at least a few hundred positive and negative examples in the test set. If your total dataset is very small (under 500 rows), consider using cross-validation entirely instead of a fixed split.