Skip to content

Training Profiles and Splits

The coverage model configs define two named profiles. Both split labels are stored in each shared model-period shard, so the large period/history dataset is built once and reused.

Profile Assignment Purpose
worker Every period for a worker follows a stable seeded hash assignment. Worker-held-out evaluation, matching the established workflow.
time Periods before 2013 are train; 2013–2016 are validation; 2017 onward are test. Time-forward evaluation.

The worker fractions and seed are configured under training_profiles.worker.splits. The date boundaries are configured under training_profiles.time.splits. Boundaries are applied to period_start: train_before is exclusive for validation, and valid_before is exclusive for test. Thus a period dated 2013-01-01 is validation, while 2017-01-01 is test.

Select a profile for each profile-aware model command:

python -m training.cross_validate --config config/config_one_percent.yaml --profile time
python -m training.train --config config/config_full.yaml --profile worker --mode training

The selected profile scopes training checkpoints, plots, evaluation outputs, cross-validation trials, score datasets, and embedding exports under its own configured roots. Shared processed employment data and model-period shards remain common. Downstream analysis pipelines are migrated separately and must be pointed at the corresponding profile outputs before comparing results.

Training modes retain the same semantics for both profiles:

Mode Fit splits Evaluation split Use
cross_validation train valid Hyperparameter and epoch selection
training train, valid test Held-out evaluation
full_training train, valid, test None Final scoring model