← Concept Index

Training & adaptation

Training data and its limits

Also called: what's in the data, data bias, where knowledge comes from

DEFINITION

A model's abilities and blind spots are shaped by the text and images it was trained on — which languages, sources, viewpoints, and time periods were and weren't well represented.

WHY IT MATTERS

It explains why models are stronger in some languages and topics than others, why they can repeat historical biases, and why they're frozen at a cutoff. The data is the model's whole world; its gaps become the model's gaps.

COMMONLY CONFUSED WITH

A neutral, complete view of the world. Training data is a partial, skewed sample, and the model inherits that skew rather than correcting for it.