nemo_gym.data.tasks.harbor.dataset_config

The [gym] table of a dataset’s dataset.toml: dataset-level settings for its tasks.

task.toml is the task author’s file and, once fetched, read-only. What a dataset needs on top of it goes here, as data: the facts Harbor has no field for (license and type, which say which split the rows may enter; allowed_agents, the harnesses admitted to run it; domain and verified, catalog metadata; source, the hub reference a committed entry under environments/harbor/ is pulled from), defaults that scale every task or set how many rollouts an evaluation takes per task, and per-task overrides expressed with Harbor’s own [environment] fields. The loader applies the task-shaping settings before anything else sees a task, so servers, agents and rows only ever see the effective task.toml; the other keys flow into the generated run config’s datasets: entry (nemo_gym.data.tasks.harbor.materialize.run_config). Harbor’s own tables in the same file ([dataset] and friends) are left to Harbor.

Deployment settings (which sandbox provider, registry credentials, image mirrors) do not belong here: a dataset must run unchanged on every deployment.

Module Contents

Classes

Name	Description
`GymDatasetConfig`	The whole `[gym]` table.
`GymDatasetDefaults`	`[gym.defaults]`: generic policy applied to every task of the dataset.
`GymTaskOverride`	`[gym.tasks."<task_id>"]`: per-task overrides, written with Harbor’s own fields.

Functions

Name	Description
`_added_artifacts`	-
`_override_environment`	-
`_override_verifier`	-
`_scale_environment`	-
`_scale_step_timeouts`	One `[[steps]]` entry with its own `[steps.agent].timeout_sec` and `[steps.verifier].timeout_sec` scaled
`apply_dataset_config`	The effective `task.toml` for `task_id`: overrides first, then the dataset defaults.
`read_dataset_config`	The `[gym]` table of `<folder>/dataset.toml`; empty when the file or the table is absent.

Data

API

class nemo_gym.data.tasks.harbor.dataset_config.GymDatasetConfig()

Bases: BaseModel

The whole [gym] table.

_path

Path | None = PrivateAttr(default=None)

allowed_agents

list[str] | None = None

defaults

GymDatasetDefaults = Field(default_factory=GymDatasetDefaults)

domain

Domain | None = None

effective_type

str

The split the rows enter: the declared one, else validation with a license, else benchmark.

is_empty

bool

Whether the [gym] table says nothing at all: every key at its default.

What a scaffolded dataset.toml reads as (its keys are comments). Nothing branches on it: a task’s digest folds the table in by shapes_tasks, and the entry and run defaults read each key.

license

DatasetLicense | None = None

model_config

= ConfigDict(extra='forbid')

path

Path | None

shapes_tasks

bool

Whether any setting changes the effective task.toml (identity and run fields never do).

source

str | None = None

tasks

dict[str, GymTaskOverride] = Field(default_factory=dict)

type

DatasetSplit | None = None

verified

bool = Field(default=False, strict=True)

nemo_gym.data.tasks.harbor.dataset_config.GymDatasetConfig._allowed_agents_name_distinct_agents(
    value: list[str] | None
) -> list[str] | None

classmethod

nemo_gym.data.tasks.harbor.dataset_config.GymDatasetConfig._domain_is_one_of_gyms(
    value: typing.Any
) -> typing.Any

classmethod

Refuse a domain outside Gym’s values with the values named: the entry’s domain is the same enum.

nemo_gym.data.tasks.harbor.dataset_config.GymDatasetConfig._source_is_a_hub_reference(
    value: str | None
) -> str | None

classmethod

Refuse a source that is not a well-formed harbor: reference, in the hub’s own words.

nemo_gym.data.tasks.harbor.dataset_config.GymDatasetConfig._training_splits_need_a_license() -> nemo_gym.data.tasks.harbor.dataset_config.GymDatasetConfig

nemo_gym.data.tasks.harbor.dataset_config.GymDatasetConfig.admits(
    agent: str
) -> bool

Whether agent (an implementation name) may run the dataset: any when allowed_agents is unset.

class nemo_gym.data.tasks.harbor.dataset_config.GymDatasetDefaults()

Bases: BaseModel

[gym.defaults]: generic policy applied to every task of the dataset.

model_config

= ConfigDict(extra='forbid')

num_repeats

int = Field(default=1, ge=1, strict=True)

resource_multiplier

float = Field(default=1.0, gt=0)

timeout_multiplier

float = Field(default=1.0, gt=0)

class nemo_gym.data.tasks.harbor.dataset_config.GymTaskOverride()

Bases: BaseModel

[gym.tasks."<task_id>"]: per-task overrides, written with Harbor’s own fields.

environment and verifier take any key of the task’s [environment] and [verifier] tables; env merges over the task’s own and every other field replaces the task’s value. artifacts are added after the task’s own artifacts entries, in Harbor’s shape (a path, or a table with source, destination and service).

artifacts

list[Any] = Field(default_factory=list)

environment

dict[str, Any] = Field(default_factory=dict)

model_config

= ConfigDict(extra='forbid')

verifier

dict[str, Any] = Field(default_factory=dict)

nemo_gym.data.tasks.harbor.dataset_config._added_artifacts(
    entries: list[typing.Any],
    task_id: str
) -> list[nemo_gym.data.tasks.harbor.models.HarborArtifact]

nemo_gym.data.tasks.harbor.dataset_config._override_environment(
    environment: nemo_gym.data.tasks.harbor.models.HarborEnvironment,
    override: dict[str, typing.Any],
    task_id: str
) -> nemo_gym.data.tasks.harbor.models.HarborEnvironment

nemo_gym.data.tasks.harbor.dataset_config._override_verifier(
    verifier: nemo_gym.data.tasks.harbor.models.HarborVerifier,
    override: dict[str, typing.Any],
    task_id: str
) -> nemo_gym.data.tasks.harbor.models.HarborVerifier

nemo_gym.data.tasks.harbor.dataset_config._scale_environment(
    environment: nemo_gym.data.tasks.harbor.models.HarborEnvironment,
    multiplier: float
) -> nemo_gym.data.tasks.harbor.models.HarborEnvironment

nemo_gym.data.tasks.harbor.dataset_config._scale_step_timeouts(
    step: typing.Any,
    multiplier: float
) -> typing.Any

One [[steps]] entry with its own [steps.agent].timeout_sec and [steps.verifier].timeout_sec scaled once; a budget the step does not name stays None, inherited through HarborTaskConfig’s step accessors (the loader’s hint about steps graded within the step default keys on that).

nemo_gym.data.tasks.harbor.dataset_config.apply_dataset_config(
    config: nemo_gym.data.tasks.harbor.models.HarborTaskConfig,
    task_id: str,
    dataset: nemo_gym.data.tasks.harbor.dataset_config.GymDatasetConfig
) -> nemo_gym.data.tasks.harbor.models.HarborTaskConfig

The effective task.toml for task_id: overrides first, then the dataset defaults.

nemo_gym.data.tasks.harbor.dataset_config.read_dataset_config(
    folder: pathlib.Path
) -> nemo_gym.data.tasks.harbor.dataset_config.GymDatasetConfig

The [gym] table of <folder>/dataset.toml; empty when the file or the table is absent.

nemo_gym.data.tasks.harbor.dataset_config.DATASET_FILE = 'dataset.toml'

nemo_gym.data.tasks.harbor.dataset_config.DatasetSplit = Literal['train', 'validation', 'example', 'benchmark']

nemo_gym.data.tasks.harbor.dataset_config.GYM_TABLE = 'gym'

nemo_gym.data.tasks.harbor.dataset_config.SCALED_RESOURCES = ('cpus', 'memory_mb', 'storage_mb')

nemo_gym.data.tasks.harbor.dataset_config.TRAINING_SPLITS = ('train', 'validation')