Concepts

Targets, deployments, instances, and datasets — the words the CLI uses

A target is what a run or evaluation uses. Targets are named by flags, and so is media: run takes it with --media-path. On a machine with an instance, a media-only run uses the instance's configured Pop.

Runnable targets

Target	Meaning	Discover it with
Model	A model alias and optional tag, such as eyepop.person:latest	eyepop get models
Configured model	A model with a prompt, configuration, and output shape; called an ability by CLI commands	eyepop get abilities
Pop	A named pipeline of one or more models	eyepop get pops
Session	A pipeline with compute and a UUID	eyepop get sessions

run --model accepts models and configured models. run --pop runs a Pop. run --session reuses an existing session.

Ways to run

Need	Use
Try a target on media	A one-off run
Keep a target warm for an application	A deployment
Run on hardware you control	An instance
Score a configured model across a dataset	An evaluation

A deployment is a session that can be reused between runs. See Persistent Deployments.

An instance is the EyePop runtime configured on one machine you control. It serves one Pop at a time. See On-Premise Instances.

Datasets

A dataset is a named collection of media. Each item is an asset and may have ground truth, the expected result used by an evaluation. A partition is a named subset such as train or test.

eyepop get datasets
eyepop evaluate --ability my-namespace.find-kittens:latest --dataset people

Next steps

• Run Inference & Evaluations — put a target and some media together

• Persistent Deployments — keep a target warm behind a stable UUID

• On-Premise Instances — run the same targets on your own hardware