Training & Alignment
Training Data
The material from which a model draws its patterns: text, images, tables, click traces. Its selection decides the capability and the slant of the result. What is missing, the model cannot do; what is skewed, it skews along with it. The industry's most expensive disputes are currently being fought over the origin and rights of this data.
An example
A model for screening job applicants is trained on ten years of hiring decisions. It learns exactly what it was meant to: the patterns in those decisions. If systematic biases were among them — certain universities favored, résumé gaps penalized, part-time work read as a signal — then they are now in the model, applied more consistently than ever. Automation does not make a practice neutral, it makes it scalable.
A closer look
For large models, they come predominantly from the open web, supplemented by licensed collections, code, and increasingly generated data. Composition and filtering shape a model's properties more than the architecture does, and are almost never disclosed.
Why it matters
Training data is the actual substance of a model, and the three questions about it are uncomfortable but short. Where does it come from, and was it allowed to be used for this? Whom does it represent and whom not? What period does it cover? For purchased models, the first question usually can't be answered. That is a risk you can know and accept but not negotiate away. With your own data it can be answered, and that is exactly why you should ask it before training, not after.
Commonly misunderstood
Whatever skews a model brings with it comes from this selection. Afterward it can be mitigated but not removed.
Where this term does its work in the magazine.
2 articles use the term or take up its principle.
From Science
Benchmark Performance ≠ Bedside Performance
A good average can hide precisely the cases in which a system becomes dangerous.Read the articleFrom Science