A model's developers need to know whether it actually improved. They measure this by running it on tasks with known answers and comparing results. If the test material was part of the model's training, the score reflects memory, not ability. Public benchmarks degrade for this reason, as their contents spread across the web and into training sets.
Private company records make strong evaluation data because they were never public. A set of real cases with recorded outcomes, such as tickets and their resolutions or claims and their decisions, lets a lab test whether a model reaches the result that a trained professional reached. Evaluation sets are usually much smaller than training sets, but they need more precise documentation: what the task is, what counts as correct and where the data has been. A license for evaluation typically forbids using the data for training, and buyers often want exclusivity so that no other model is trained on the same material.
A seller can raise the value of an evaluation set by documenting the outcome of each case clearly and by keeping a written record showing the data has not been shared elsewhere.