Each stage establishes the inputs for the next. Published datasets and checkpoints let users enter this workflow at training or evaluation.
| Stage | Transformation and responsible code | Check before proceeding |
|---|---|---|
| Simulation | Detector-specific generation under mlpf/data/ produces ROOT detector signals and truth relationships | Expected collections, events, and production configuration |
| Postprocessing | CMS, Key4HEP, or ColliderML converters construct element features and assigned targets in pickle or Parquet | Schema, associations, closure, and geometry checks supported by the converter |
| Dataset construction | Builders under mlpf/heptfds/ package a representation into versioned TFDS splits | Decode events; confirm names, versions, feature shapes, and split contents |
| Training input | PFDataset transforms momentum targets, batches variable-size events, and optionally constructs particle sets | Matching model recipe, feature schema, masks, and target counts |
| Training | mlpf/pipeline.py train optimizes the model and saves checkpoints and configuration | Finite losses, held-out validation, and complete experiment artifacts |
| Inference | The standalone Key4HEP evaluator or detector-specific validation workflow decodes predicted particles | Correct checkpoint/configuration pairing and readable output collections |
| Validation | Particle, jet, missing-momentum, and export tools compare outputs with reference quantities | Physics response and, for deployment, PyTorch/runtime numerical agreement |
The input and target guide explains the data carried through these stages; the architecture guide explains the learned transformation.
For commands, follow dataset generation, dataset validation, training, and model validation. Keep the checkpoint, saved model configuration, dataset version, and software revision together as described in the artifact reference.