A model’s training examples reflect choices about collection, selection, and representation. For a supervised task, labels also define the distinctions the model is being asked to learn.
Consider whether the examples cover the situations in which the system will be used. A collection can be large while leaving important cases underrepresented. Inconsistent labels can make the intended task unclear even when the inputs themselves are plentiful.
When evaluating a system, ask how the task and data relate to the real use case. The amount of data is one part of the story; its relevance, quality, and coverage are others. Those questions help connect model performance with the work it is expected to support.
An example to consider.
Consider a collection of sample labels that covers only easy cases. An unfamiliar input may expose a boundary the examples never explained.
Put it in perspective.
Give review a specific responsibility. Checking a source, testing an instruction, and editing a sentence are different jobs and may need different evidence.
- Ask which situations the examples cover.
- Look for a clear labeling rule.
- Connect the training task with the intended use.
Follow a related question
Identify unnecessary transfers or repeated work.
Build with a little more restraintTry a small offline task.
When a tool keeps working offlineKeep learning
Related background to continue exploring this subject.
Google: an introduction to language models NIST: AI risk management framework
