Glossary - TunePlane

These words appear in error messages. Reaching for a near synonym — run for job, image for recipe, gym for environment — usually lands you on the wrong page.

Your project

Term	Meaning
Project	The `name` in `tuneplane.yaml`. The console groups every job from this directory under it. Names are unique per user, not globally.
Experiment	A directory under `experiments/<name>/`, holding `config.yaml`, `recipe.lock.json`, and for a custom job `train.sh`. Creating one starts no GPUs.
Job	One admitted submission. It has a run id, a state, and once launched a backend handle. One experiment produces many jobs.
Run id	The identity everything else refers to: artifacts, `--init-from`, share links.
Pipeline	One submission and every Step the platform runs for it. The unit the job list lists.
Step	One job inside a Pipeline, with exactly one Step Role: `train`, `export`, `evaluate`, `benchmark`, `serve` or `process`.
Root Step	The Step you submitted. The only one that can carry a Derivation or belong to a sweep.
Derivation	That a run began from another run’s artifact. Crosses projects and owners; Pipeline membership never does.

Methods and packaging

Term	Meaning
Method id	`<framework>/<method>`, for example `nemo-rl/grpo` or `trl/kto`.
Recipe	The versioned declaration of a method: entrypoint, tunable parameters with ranges, metrics, artifacts, and which framework versions it runs on.
Method Category	What kind of training a recipe is, from a closed set: `supervised`, `sft`, `preference`, `reward-model`, `rl`, `distillation`, `evaluation`, `custom`. The axis is what the method needs in order to run.
Catalogue	The recipes the connected server has published. Your lock file and the server must agree.
`recipe.lock.json`	Written by `tp new`. Pins the recipe digest and the framework version. Submit refuses a stale lock — upgrade with `tp recipe upgrade`.
JobSpec	The JSON `tp` builds and the server admits. You do not hand-write it. See JobSpec.
Job Capsule	`capsule.json`, `bootstrap.sh` and `runner.pex`, injected by the server under `.tuneplane/capsule/`. The training image does not have the platform package installed.
`runtime_id`	Which deployment artifact runs the job — an OCI image, a SIF, a SQSH. Resolved from `--image`, the framework default, the runtime registry, then the catalogue.

Hardware and placement

Term	Meaning
Series	One class of accelerator, as the deployment defines it: per-card memory, compute weight, the scheduling pin, the Slurm constraint, and how many cards exist. Quota and capacity are per series.
Profile	A named preset over a series: a default shape plus env and framework overrides for that shape. `h200`, `h200-2g`, `h100`. What `--profile` takes.
Resource Pool	A group of same-series cards within one job, which the job’s roles are assigned to. A shape in the JobSpec, never a set of machines.
Fleet	A named execution target: registered Nodes or an existing Kubernetes or Slurm cluster. Jobs are placed onto a Fleet. Its kind is `node`, `kubernetes` or `slurm`.
Node	One registered machine in a Fleet. Its id does not survive a reinstall: a rebuilt machine is a new Node.
Executor	How the control plane starts a container on a Fleet. One per Fleet kind, with no fallback between them.

Data and material

Term	Meaning
Dataset	Versioned training material, referenced `<owner>/<name>[@version]`. A version is immutable.
Volume	A named directory a job reads through a read-only mount. It has no versions: files change in place, and a reference names the volume. Downloading its files is off until the owner turns it on.
Hub Reference	`[<hub>:]<namespace>/<name>[@<revision>]`. Unprefixed means this platform; `hf:` and `ms:` mean Hugging Face and ModelScope.
Hub Import	Copying a hub repository once into this platform’s registry, so preview, scanning, redaction, classification and quota apply to it.
Data Classification	The handling policy a derived dataset inherits from the material it was produced from. Computed by the platform, not declared by the job.
`$TUNEPLANE_OUT_DIR`	The one directory whose contents survive the job. A checkpoint written anywhere else disappears with the scratch tree.
Ingest	The HTTP path from the training container back to the console: logs, metrics, samples, artifacts. Its address arrives in the job as `TUNEPLANE_JOB_ENDPOINT`, and it must be reachable from the GPU node.

Evaluation

Term	Meaning
Benchmark pack	A declaration of an evaluation set: image, argv, and how to map its output onto score fields. Runs through `tp bench run`.
Rubric	Your team’s written standard for a correct answer: named criteria, weights, a scale. Owned and versioned, referenced `<owner>/<name>`.
Environment	A taskset plus a manifest, owned and versioned like a dataset. What an agent practises against.
Taskset	The tasks in one environment version. Never handed out: a job references one, and no route returns a copy.
Verifier	What decides whether a task was completed — a Rubric, a plugin entrypoint, or an HTTP endpoint. Always a reference; the platform never is one.
Harness	The code that drives a model through an environment, supplied as a `capability: environment` plugin.
Sandbox	Where model-generated code runs: a throwaway container with no GPU, no network and hard limits. A job that needs one on a deployment without one is refused.

Serving

Term	Meaning
Playground Session	A temporary serving instance for a human to try a checkpoint. Not a Deployment.
Model Deployment	A managed model version for application traffic, behind a stable address, with a declared desired state.
Deployment Revision	An immutable snapshot of a Deployment’s model source and serving configuration. Promoted or rolled back whole.
Deployment Token	A revocable credential for application traffic to one Deployment, independent of anyone’s login.
Serving Engine	The runtime that answers inference requests. vLLM and SGLang are different engines.
Reflow Buffer	The retained request and response records of one Deployment, under a retention clock. Not a dataset: it expires, and nothing has vouched for it.

Next

Job states · Resources · Architecture