Training data

Training data is the collection of examples a model learns from. Its size matters less than its representativeness: a model trained on examples that do not cover a situation will behave unpredictably the first time that situation occurs in production.

It is also the main legal and ethical surface of a machine learning project. Whether the data may lawfully be used for training, whether it contains personal data, whether it encodes a historical bias you are about to automate — all of these are questions about the data, not the algorithm.

For custom models, assembling and labelling training data is usually the largest line in the estimate, and the one most often underestimated. It is worth pricing separately and honestly rather than folding it into "model development".