LoopArtifacts
Per-run filesystem + state pointers, used to populate RunResult.
Attributes
attributerun_dirPathThe output directory written to. Algorithms surface this as
RunResult.artifacts["run_dir"] for downstream consumers
(e.g. TinkerInference.from_run_result).
attributemanifest_pathPathPath to the checkpoints.jsonl the loop wrote.
attributecheckpointslist[ManifestRow]Manifest rows the loop wrote, in order. The last one is the final sampler - the URI eval consumes.
attributetotal_requested_stepsintThe step horizon passed to :meth:TrainingLoop.run (num_steps) -
NOT the number of steps actually executed. These differ when a callback
early-stops the loop via state.request_stop(): this field still reads
the requested ceiling. For "how many steps actually ran", read
state.step + 1 inside on_train_end (state.step is the last
executed index).
attributetrain_secondsfloatFunctions
funcas_dict(self) -> dict[str, str]Flatten to the RunResult.artifacts shape that downstream
consumers (e.g. TinkerInference.from_run_result) read.
paramselfReturns
dict[str, str]func__init__(self, run_dir, manifest_path, checkpoints, total_requested_steps, train_seconds) -> Noneparamselfparamrun_dirPathparammanifest_pathPathparamcheckpointslist[ManifestRow]paramtotal_requested_stepsintparamtrain_secondsfloatReturns
None