Training data

Training data is the textbook a model studies: thousands (or billions) of examples, each usually paired with the right answer — this email is spam, this photo is a cat. The model tunes itself until its answers match the labels.

Garbage in, garbage out: a model trained on biased, sloppy, or unrepresentative examples faithfully learns those flaws. Serious ML work is often 80% caring for the data and 20% everything else.