Glossary

Short definitions of the terms used throughout the portal. For how they fit together, see Core concepts.

Workspace : The top-level container for a team or account. Holds projects, members, and API keys.

Project : A named grouping of related runs — typically one model, dataset, or research effort.

Run : A single training session (one execution of your code). The central object you analyze; it has attributes and metrics.

Attribute : A typed key-value piece of run metadata (string, number, boolean, date, or set of strings) — for example a hyperparameter from your config. Used for table columns and filters.

Metric : A value your code logs repeatedly over a run. In the portal a metric is plotted as a series.

Series : A metric’s time-series, identified by its name, scale, and labels. See Metrics, series & attributes.

Scale : The x-axis dimension a series is indexed against — step, RL step (ML_STEP), episode, or environment step (ENV_STEP).

Label : An optional key/value dimension on a series (e.g. layer="3") that distinguishes variants of the same metric.

Series summary : Precomputed statistics for a series (first, last, min, max, mean, …) used for fast displays such as table columns.

Environment (reinforcement learning) : A single instance of the world an agent interacts with during an RL run, identified by an env_id, producing episodes and per-step data. A run can have many running in parallel.

Episode (reinforcement learning) : One episode within an environment, indexed by an episode number; episode-level metrics are summarized per episode.

Smoothing : A chart option that reduces noise in a series for readability (e.g. EMA, moving average, gaussian).

Inherited metrics : Metrics that come from higher in the project hierarchy, optionally included in a view via the include inherited chart setting.