Checkpoint Runner API

All-seed preflight compares saved constructor hyperparameters in both last and selected checkpoints against the current arm config: LoRA rank/alpha/dropout and, under active MoE, expert count/top-k/hidden dimension/load-balancing coefficient. Missing required controls fail closed; the arm retains its 21 settings keys.

All-seed preflight validates complete monitor epochs, settings and the unambiguous earliest best checkpoint before exports or decision reads. The returned artifacts carry the monitor records.

Read trained seeds through their selected checkpoints (spec 4.6).

check verifies the durable monitor records and both checkpoints without loading a model or exporting a table. tools sweep checks every seed before its first decision read. run then exports the selected checkpoint's table and returns the encoder and the reads that selected it.

CheckpointRunner

Load trained seeds from their run directories, refusing incomplete or inconsistent runs.

Parameters:

Name Type Description Default
cfg Config

The arm's configuration, including the descriptions and tokenization window.

 required
bundle ValidatedSupervisionBundle

Its validated supervision bundle.

 required
run_directory Callable[[int], Path]

Resolve each seed's checkpoint directory, as pulled from training.

 required
device Union[str, device]

Where exports and query encoding run; independent of the training accelerator.

 required

check(spec, seed)

Check one seed without exporting or reading a decision panel (P25, spec 5).

Records must cover every epoch through last.ckpt exactly once, name its training run and read under the distance of the arm's geometry (Req 12). The earliest epoch with the highest MRR must have one unambiguous checkpoint, whose identity, contract, settings and saved best score agree exactly.

Raises:

Type Description
ValueError

Naming the seed and its first missing or inconsistent artifact.

run(spec, seed)

Export the checked selected epoch, replacing any existing table and provenance pair.

SeedSelection dataclass

A seed's earliest best epoch, its checkpoint and the monitor records through last.ckpt.