Definition
Training data comprises the examples and signals used to fit a model's learned parameters. It can include raw observations, labels, demonstrations, preferences, rewards, synthetic examples, and transformed or filtered derivatives used during pretraining, fine-tuning, or reinforcement learning.
Training data should be described by provenance, collection and consent, licensing, filtering, deduplication, labeling, coverage, representativeness, contamination, retention, and version, not merely by record count.
Distinguish it from nearby terms
A dataset can support training, validation, testing, or evaluation. Only the portions that influence parameter fitting are training data. Runtime prompts, retrieval corpora, and production feedback are not training data unless a later learning process uses them to update the model.
Check your understanding
Ask whether an example changed model parameters, selected model settings, evaluated performance, or only supplied runtime context; those are different data roles.