Metrics

The metrics service computes evaluation metrics during validation and testing. It holds a user-selected list of metrics, updates them as batches are evaluated, and reports their values. It is one way a run can measure quality beyond the training loss.

Usage

Configured under services.metrics. The select list names the metrics to compute, each a selectable plugin.

services:
  metrics:
    select:
      - name: top-k-acc
        kwargs: { k: 1 }

How it works

Each selected metric accumulates state incrementally as batches arrive and resolves to a per-head value when computed, so metrics are exact over the whole evaluation set and combine correctly across processes under distributed execution. The individual metrics are documented under the metric slot.

Configuration

Option	Description	Type	Default
`select`	Ordered list of metric selections, each a `name` / `kwargs` pair.	list[mapping]	required

Sub-slots

Metric