Feature Matrix

Usually written $X$, with shape $n \times p$ for $n$ examples and $p$ features. Statisticians call it the design matrix. The target sits alongside it as a separate vector, normally $y$.

It's the model's entire universe. Whatever is in $X$ can be learned from; whatever isn't, can't — no matter how obviously relevant it is or how carefully it's stored elsewhere.

One shape detail catches people out: the row count is examples and the column count is features, so a matrix with more columns than rows means more parameters than data points, and an unregularised linear model can then fit any labels exactly.