Concepts

Note

This page will fill in as the API surface stabilizes. Track progress in the Linear Orion project.

Topics planned:

• RunConfig and its sub-models (TrainingConfig, DataConfig, ModelConfig, InfraConfig, BenchCompute)

• OrionTrainer and the standard callback set

• The orion.models.* wrapper pattern around torch_brain

• Rich checkpoints — what’s in them, how resume works

• The data_hashes/manifest.json and train/test overlap detection

• The Benchmark class as a first-class measurement primitive

• BenchCompute modes and partial subsets

• Multi-stage training with TrainingPipeline and FinetuneFrom

• Lineage propagation

• Pre-flight validation

• Module version routing for orion.models.*