Chapter 1 ended on a checklist, and one of its three questions — are there recorded examples? — deserves a chapter to itself. This is it. Where training data actually comes from, how the data team reads a table of past cases, why so much of it arrives broken, and the oldest law of the field: a model can only be as good as its examples.
5 topics
Conversations about machine learning gravitate to the model — the training, the clever algorithm, the demo. Data gets a supporting role, mentioned once and skipped. Practice runs the other way around: the data team spends most of its working life on the examples, because the examples decide what any model can ever know. The model is the student; this chapter is about the textbook.
Five topics build the picture. First, the training table itself — rows, columns, and the one column that holds the answer. Then the kinds of data beyond tidy tables — photos, text, sound — and the single trick that lets models eat all of them. Then the sources: who actually produces two million rows, and who gets paid when reality doesn't hand over the answers. Then the mess — the holes, duplicates, and impossible entries that fill every real dataset. And finally the law that governs everything downstream, sharp edge included: garbage in, garbage out.
From the world to a model — the path every training dataset travels