Configuration¶
The model configs are coverage-specific:
| Coverage | Model config |
|---|---|
| Full | model/config/config_full.yaml |
| One percent | model/config/config_one_percent.yaml |
| Samples | model/config/config_samples.yaml |
The key sections are:
| Section | Controls |
|---|---|
processed |
Inclusive employment observation window |
data |
Processed inputs and model-period datasets |
training_features |
Worker, firm, history, and relationship features |
training_profiles |
Worker/time split definitions, artifact roots, and selected training checkpoints for scoring |
full_training_artifacts |
Shared full-training roots and selected full-training scoring checkpoint |
model |
Tower architecture, hazard baseline, and optional heads |
training |
Optimization, batching, device, and loader settings |
training_modes |
Which splits each fitting mode uses and evaluates |
cross_validation_search |
One-factor trial definitions |
scoring |
Scoring runtime and dataset-specific options; paths are derived from the selected profile/root |
processed.observation_end_date must match the corresponding shared coverage
config. Together with the model-only processed.observation_start_date, it
defines the model-period observation window. The start date excludes earlier
period rows while retaining earlier employment history to calculate tenure and
worker history; the end date is also applied upstream for right-censoring.
Each config defines worker and time profiles. The model-period builder
stores both split labels in one shared set of period shards. Select one with
--profile worker or --profile time when running cross-validation, training,
benchmarking, scoring, or embedding export. Training outputs are isolated under
the profile's run_root; score and embedding outputs use its score_root.
For profile-aware training scoring, set
training_profiles.<profile>.scoring_checkpoints.training to the selected
checkpoint. For profile-independent full_training, set
full_training_artifacts.scoring_checkpoint. Scoring output directories are
derived from the selected score_root, mode, and dataset; do not duplicate
those paths under scoring.