Configuration
Portable templates
The JSON templates in configs/examples/ contain the scientific training settings and relative paths.
| Core | Correction interval | Template |
|---|---|---|
| Explicit | 1 hour | explicit_1h.json |
| Explicit | 1 day | explicit_1day.json |
| Semi-implicit | 1 hour | implicit_1h.json |
| Semi-implicit | 1 day | implicit_1day.json |
evaluation_100d.json describes paired frozen forecasts. rl_posttrain.json demonstrates the optional score-function trainer described in Posttrain.
Training fields
| Field | Meaning |
|---|---|
core | explicit or implicit_p0_fixed |
frequency | 1h or 1day correction interval |
data_root | Matching reference ensemble |
run_root | Caches, calibration, weights and run records |
world_size | Number of distributed training ranks |
grid | Coarse horizontal dimensions |
model.hidden_dim | CorrectorNet width |
split | Members, model-day bounds and sample strides |
stages | Stage budgets, rollout horizons and checkpoint intervals |
losses | Loss names, weights and normalization scales |
acceleration | Physical execution and solver backends |
wandb | Project, entity and online or offline mode |
Training paths resolve relative to the configuration file. Defaults and their merge rules are implemented in src/neuralpom/v1/config.py. The plan command prints resolved budgets:
python scripts/v1/plan.py --config configs/local/explicit_1day.jsonThe portable examples select offline W&B. Set wandb.entity to your own account or team and wandb.project to your project when enabling online tracking.
Evaluation protocol
The paired evaluator uses a separate protocol with groups, cases, horizons_days, correction_policies and output_root. Each group names a core, correction interval and training configuration. Each case names a reference member and initial model day. The worker's --variant selects one group, stage and correction policy; the quick start shows a complete command.
The training evaluator in scripts/v1/evaluate.py uses canonical constrained correction. The campaign evaluator in neuralpom.evaluation.campaign additionally exposes raw, constrained and density75 policies.