Metrics, series & attributes

This page explains the data inside a run: metrics (the time-series you plot) and attributes (the metadata you filter and group by). A clear picture of these makes charts, filtering, and comparison much easier to reason about.

What identifies a series

A metric is logged as a series — a sequence of numeric values over the course of a run. A series is identified by three things:

• Name — what you logged, e.g. loss or train/accuracy.
• Scale — the x-axis dimension the values are indexed against (see below).
• Labels — optional key/value dimensions attached to the values (e.g. layer="3"), used to distinguish variants of the same metric.

Two values with the same name but different scales or labels belong to different series. This is why the metrics tree may show several entries under one familiar name.

Scales

The scale determines what a point’s x-position means:

• Step — the iteration counter for standard (supervised) training; increments per logged value.
• RL step (ML_STEP) — a reinforcement learning gradient-update counter.
• Episode — an RL episode boundary.
• Environment step (ENV_STEP) — a per-step counter within an RL episode.

In charts, you can additionally view any series against relative time (elapsed since the run started) or absolute time (wall-clock) — see X-axis mode. Reinforcement learning scales are covered in RL concepts.

Series summaries

Alongside the raw points, Metrana keeps a summary of each series — first, last, min, max, mean, and similar statistics. Summaries power fast displays such as a metric’s value as a column in the runs table, without loading every point.

Attributes vs metrics

Both describe a run, but they behave differently:

	Attribute	Metric (series)
Shape	A single typed value	A time-series of values
Examples	learning_rate, model, created, tags	loss, reward, accuracy
Where you see it	A run’s Overview; table columns/filters	Charts, Dashboard, Compare
Typical use	Filter, sort, and group runs	Plot and compare trends

Attributes are typed — string, number, boolean, date, or a set of strings — which determines the operators available when you filter a table and the icon shown next to each field.

Field sources

When you pick fields to show as table columns or to filter by, they come from a few sources:

• Default — built-in run fields (name, …).
• Attribute — metadata your code logged via config.
• Series summary — a statistic derived from a metric series (e.g. the last value of loss).

Knowing the source helps explain why some columns are single values (attributes, summaries) while charts work from the full series.

Related

• Core concepts — how runs, projects, and workspaces fit together.
• Chart settings — scales and smoothing applied when plotting.
• The runs table — filtering and columns built on attributes and summaries.