Model outputs¶
Model outputs are datasets written by an HCT scorer. training outputs are
profile-specific: worker holds out workers, while time holds out future
periods. They use separate checkpoints and output roots. full_training is
profile-independent, uses all configured splits, and provides the final MQ
measure used by most downstream analyses.
The model-output workflow gives the ordered scoring commands for the reusable observed-match score output. Analysis-specific evaluation data is built subsequently in the analysis-data workflow.
- Worker-firm-period scores contain MQ and predicted separation hazards for observed worker–firm periods.
- Embedding tables provide reusable company vectors, mover-only worker-state vectors, and vectors for every worker relationship-period.
Training and architecture are documented under Model.