EvSys

Evaluator

Periodic in-loop evaluator.

Each evaluator owns one metric source (a Benchmark, a custom probe, etc.) and carries its own run_every cadence. This lets a single training run mix a fast val benchmark (run_every: 50) with a heavier test benchmark (run_every: 500).

Attributes

attributenamestr

Short name; used to prefix the metric keys (val/\<name>/\<key>).

attributerun_everyint

Per-evaluator step cadence. 0 → disabled (never fires). Positive → fire when (step + 1) % run_every == 0.

Functions

funcevaluate(self, sampler, *, model_path=None, step=None) -> dict[str, float]
paramself
paramsamplerSamplingClient
parammodel_pathstr | None
= None
paramstepint | None
= None

Returns

dict[str, float]

On this page