Knowledge
A shared language for better data.
A short guide to the ideas behind training, evaluation, and useful AI behavior.
Supervised fine-tuning
Training a model with examples of a desired response or behavior for a given input. The examples establish the task and the standard to aim for.
Preference data
Comparisons between possible responses, made against stated criteria. They describe which alternative better serves the task and why.
Evaluation
A defined set of tasks and measures used to assess a system. A useful evaluation makes the required behavior and the basis for grading explicit.
Rubric
A set of criteria that guides assessment. It can separate dimensions such as factual support, instruction following, completeness, and clarity.
Agent environment
A setting in which a system can observe information, use tools, and change a state. Tasks can be assessed by checking the resulting state.
Computer-use trajectory
A sequence of observations and actions through a browser or desktop task. The surrounding context and completion conditions help make the sequence interpretable.
Data provenance
A record of where data came from, how it was prepared, and the rights or restrictions attached to its use.
Held-out tasks
Tasks kept separate from the material used for development or training, to provide a more independent view of performance.