Workflow & Validation

Seed Set

Definition

The initial set of coded documents used to begin training a TAR model.

Why it matters

Its composition shapes early model behavior and is sometimes negotiated between parties.

In practice

The seed set is the first batch of coded documents a TAR model learns from. It can come from keyword hits, documents the team already knows about, a random sample, or a mix.

In TAR 1.0, the seed set shapes early training and has sometimes been negotiated or disclosed between parties. In continuous active learning it matters less, because the model keeps learning from every document reviewed afterwards.

A seed set built only from keyword hits can steer early review toward documents that look like those hits, so many teams include some random or diverse documents.

Compare with

  • Training Set: The seed set is the start of the training set; the training set includes every document used for training.
  • Control Set: Control documents are held out from training; seed documents are the first training input.

Read more

← Back to glossary