Workflow & Validation

Training Set

Definition

The documents whose human coding decisions are used to teach the model what is responsive.

Why it matters

How it is built and documented is a frequent point of scrutiny.

In practice

The training set is every document whose human coding decision was used to teach the model. In TAR 1.0 it is built in rounds before classification; in continuous active learning, every reviewed document joins it.

Coding quality matters more than volume. Inconsistent decisions teach the model inconsistency, which is why teams track overturn rates and settle hard calls early.

Never measure the model on its own training documents. Performance on documents it has already learned from overstates how it will do on the rest of the collection.

Compare with

  • Control Set: The control set is held out from training so it can measure the model fairly.
  • Seed Set: The seed set is only the first part of the training set.
← Back to glossary