Training Data

Why You Need a Testing Set

Testing a model on the same data it learned from is like grading your own exam — you will always get 100%, but you have learned nothing.

Training data is the collection of examples a model learns from. Everything the model knows comes from it, and nothing else.

This makes it the ceiling on quality. A model trained on biased data is biased. Trained on wrong labels, it learns the wrong thing confidently. No amount of clever architecture fixes bad data.

Think of it like this. Think of a chef who has only ever eaten at one restaurant. They can cook brilliantly in that style and nothing else, and they have no way of knowing what they never tasted.

How it works

• Rows are examples, columns are features.
• Supervised data carries labels, the correct answer for each example.
• The model fits itself to these examples, adjusting parameters until predictions match.
• Data gets split. Training to learn from, validation to tune against, test to measure honestly.
• The test set is touched once. Peek at it repeatedly and it quietly becomes a second training set.

Common mistakes

• "More data fixes a bad model." More of the same bias gives you a more confident bias.
• "Labels are correct." Human labelling is inconsistent, and errors put a hard ceiling on accuracy that no model can exceed.
• "Training and test data are interchangeable." Let them overlap and your evaluation becomes meaningless. This is data leakage, and it is common.
• "Every row is worth the same." Duplicates, outliers, and stale examples can each do more harm than good.

Trade-offs

• Quantity against quality. A smaller, carefully labelled set frequently beats a larger noisy one.
• Breadth against focus. Broad data generalises; narrow data performs better on the narrow case and fails outside it.
• Freshness against stability. Recent data reflects current reality and gives you less of it.