Dataset Size (D)

The number of training tokens (or training examples). For language models, this is usually measured in tokens (millions to trillions). More data generally improves performance but requires more compute.