data

Definition

The raw material a model learns from and is measured against. Modern practice treats curation — filtering, deduplication, mixture weights, licensing — as at least as decisive as architecture, and many reported gains trace back to data rather than modeling. Keep training and evaluation data strictly separate; overlap between them, called contamination, silently inflates benchmark scores.

Explainers using this term

• Paper Walkthrough — RoboTok: Mining the Web for Demonstrations That Move Like YoursRoboTok: An Internet-Scale Data Engine for Human Demonstration Retrieval and Dexterous Manipulation Learning
• SolarWM: Training on 5 Seconds, Walking for an Hour — and Opening the Whole StackSolarWM: Open Data and Scalable Training for Long-Horizon Video World Models
• Paper Explained: What Makes Good Agentic Data? The ACE LensWhat Makes Good Agentic Data? An ACE Lens on Data Generation for LLM Agents

In the textbook

• Moving Data Costs More Than Computing