训练数据

The massive corpora of text, images and more used to train models.

The scale and quality of training data directly shape model capability: the knowledge cutoff comes from data dates, and bias and copyright issues trace back to it. Frontier models now train on tens of trillions of tokens.