Enter your dataset size and split ratios to get exact sample counts for each set.
Quick presets:
The training set is what the model actually learns from, adjusting its weights to minimize error on this data. The validation set is used during development to tune hyperparameters and make architecture decisions, checking performance on data the model hasn't trained on, but which you still look at repeatedly during iteration. The test set is used exactly once, at the very end, to get an honest, unbiased estimate of how the model will perform on genuinely unseen real-world data.
This is a common point of confusion. If you only had a train/test split and used the test set to guide every decision, you would gradually overfit to the test set indirectly, your choices would start reflecting what happens to work well on that specific test data, not genuine generalization. The validation set absorbs this development-time evaluation, keeping the test set clean and reserved for one final, trustworthy performance check.
70/15/15 and 80/10/10 are the most common starting points for small to medium datasets. For very large datasets, even a 5% validation and test split represents plenty of examples for reliable evaluation, so ratios like 90/5/5 become reasonable, since the absolute count of validation and test samples matters more than the percentage once you're working at scale. For genuinely small datasets, consider cross-validation instead of a fixed split, since a fixed 15% test set might only be a few dozen examples, too few for a statistically reliable performance estimate.
Simply splitting randomly isn't always correct. If your data has natural groupings (multiple rows from the same patient, multiple images from the same video, time-series data), a random split can leak information, the model effectively sees data from the same source in both train and test, inflating apparent performance in a way that won't hold up on genuinely new data. In these cases, split by group or by time rather than by individual row.