Checkpoint format¶

Maturity: verified

The checkpoint format defines the serialized training state and identity needed to validate and restore a completed optimizer step.

Checkpoint file¶

Checkpoint filename: checkpoints/step_<eight-digit-global-step>.pt.

Exact payload¶

metadata
training_module_state_dict
optimizer_state_dict

Unknown or missing top-level fields fail restore. Model and optimizer values are native PyTorch state dictionaries.

Metadata fields¶

Field	Contract
`global_step`	positive count of completed optimizer updates
`config_hash`	SHA-256 of the complete resolved application config
`training_contract_hash`	SHA-256 of resume-compatible training semantics
`run_id`	logical run identity and safe path component
`attempt_id`	producing attempt identity and safe path component
`model_implementation`	selected model catalog key
`objective_implementation`	selected objective catalog key
`optimizer_implementation`	selected optimizer catalog key
`dataset_fingerprint`	SHA-256 identity of the consumed canonical split
`parent_checkpoint_uri`	exact parent URI, or null for a fresh attempt

The metadata mapping is exact: unknown or missing fields fail validation. Resume validates its compatibility subset before mutating the newly constructed training module or optimizer.

Unsupported resume state¶

No scheduler, mixed-precision scaler, sampler, dataloader cursor, random-number- generator, epoch, validation, best-model, distributed-rank, or sharded state is stored.

The terminal attempt manifest, not the PyTorch payload, records the checkpoint URI, byte size, and SHA-256 content hash.