Training data
What is Training data?
The examples a model learns from before it's ever put to use. Their quality and coverage set the ceiling on what it can do: gaps, bias or errors in the data show up later in the answers. A model knows nothing beyond what its training data, plus any later updates, gave it.
Why it matters
A model knows only what its training data taught it. The quality and coverage of that data set the ceiling on what the model can do, so gaps, bias or errors in the data show up later as gaps, bias or errors in the answers. This is the “garbage in, garbage out” problem, and with AI it is amplified because the data is so vast that nobody reviews all of it by hand. For a business the risk is concrete. A hiring model trained on biased past decisions will repeat the bias, and a model trained on data that stops at a certain date knows nothing after it unless you add more.
In practice
A company builds a customer-service model on its own past chat logs and finds it inherits every bad habit in those logs, including the outdated policies staff used to give. The fix is upstream, in the data: clean it, fill the gaps, check it for the biases you care about. A model is only ever as good as what it learned from, plus whatever later updates or retrieval you bolt on.
