Module dataset 

Dataset manifests, rustev.dataset/1 (spec 004, 3.5.1).

A manifest names each case once, with its source and snapshot identity, one split, a label or an explicit missing label, and its subgroups. Its identity is content-derived. Training, model-selection, calibration and final-test membership are disjoint by construction, and a source or snapshot identity repeated across splits is refused as leakage. Semantic duplicates the manifest cannot identify (two sources with the same content) remain a stated limit: Rustev does not detect them.

rustev.dataset/2 (spec 041, 3.1.1) is version 1 with an optional gold record per case. A version 1 manifest is a version 2 manifest whose every case has no gold, and keeps its bytes and identity; each version’s identity is tagged with its own schema.

Structs

AdapterRef

A task adapter, by name and version.

AdmissionRecord

The record that admits one dataset manifest as qualification input (spec 023, 3.1). The owner writes and reviews it before any case is evaluated; Rustev checks that it is complete and agrees with the manifest, never that its statements are true. Its digest is the admission identity a qualification and every claim name. It holds no binary64 number: an agreement statistic is a Decimal string. The record and the data stay with the host, outside this repository (spec 023, 6.4).

Admitted

A manifest with a checked admission record. Only admit makes one, so holding it is the statement that the checks of 3.1 ran.

AgreementStatistic

An inter-annotator agreement statistic with its denominator.

DatasetCase

DatasetManifest

Exposure

Contamination and exposure (3.2).

LabelMapping

LabelingMethod

How labels were produced (3.1.4).

LicenseTerms

The license and what it permits (3.1.2).

ModelRef

A model, by family and exact identity.

OwnerReview

The owner’s review of this record.

Population

Who and what the data represents (3.1.7). A claim may name this population and no other (3.5).

PrivacyBasis

The privacy basis (3.1.3).

SliceSize

SplitAssignment

How cases were assigned to splits (3.1.6). Disjointness by source and snapshot is the manifest’s own check (spec 004, 3.5.1); the counts here must equal the manifest’s.

TaskMapping

How the dataset’s labels map to the task adapter’s (3.1.5).

Enums

AdmissionError

Why a dataset was not admitted.

AdmissionSource

Where the data came from (3.1.1).

DatasetError

Why a manifest was refused.

Label

A label, or an explicit statement that there is none.

LabelProducer

A label producer. A model under test, or a model of the same family, is never an independent labeler.

Constants

DATASET

DATASET_ADMISSION

Schema of an owner-reviewed dataset admission record (spec 023, 3.1).

DATASET_V2

Version 2: per-case gold records (spec 041, 3.1.1).

Functions

admit

Admit a manifest under an owner-reviewed record (spec 023, 3.1, 3.2). The record must be complete and agree with the manifest; labels must be independent of every system under test; data derived from real user data needs the owner’s privacy decision. A synthetic manifest is never admitted.