Talk to usRequest a dataset ↗
Services↗Industries↗Datasets↗Security↗Insights↗Request a dataset
Dataset documentation

Delivery should be inspectable after the files arrive.

TrainLayer pairs dataset files with documentation that explains what was built, how quality was assessed, what changed and where limitations remain.

01

Dataset card

  • Dataset name, owner and release identifier
  • Intended use and explicitly excluded uses
  • Source categories and collection period
  • Population, geography, language and modality coverage
  • Annotation schema and reviewer profile
  • Rights, consent and licensing context
  • PII or sensitive-data treatment
  • Known limitations and material risks
02

Quality report

  • Acceptance criteria defined before scale
  • Sampling strategy and sample size
  • Gold-set or calibration performance where applicable
  • Inter-annotator agreement or task-specific consistency metrics
  • Error taxonomy and defect distribution
  • Rework performed and residual open issues
  • Coverage, imbalance, duplication and leakage checks
  • Reviewer or domain-expert escalation summary
03

Version manifest

  • Semantic or project-specific release number
  • Creation and release dates
  • File inventory and schema version
  • Record counts and exclusions
  • Checksums where the delivery method supports them
  • Changes from the previous release
  • Known issues and compatibility notes
  • Correction and deprecation history
Evidence, not decoration

Metrics must be interpretable.

A single accuracy percentage rarely explains dataset quality. Reports should define the unit measured, sampling method, confidence limitations, error severity, reviewer process and whether results represent the full dataset or an audited subset.

Targets shown in proposals or sample specifications are not represented as achieved results until measured on the applicable project delivery.

Example release record

A concise version trail.

v1.0.0Initial accepted deliverySchema 1.0 · baseline dataset card and QA report
v1.0.1Patch correctionCorrected audited label defects; no schema change
v1.1.0Minor expansionAdded approved edge-case cohort and updated coverage notes
v2.0.0Material revisionNew schema or changed compatibility requiring migration
Define acceptance early

Agree on the evidence before production.

Documentation requirements, quality measures and release conventions should be part of project scoping—not decided after annotation ends.

Scope a governed dataset ↗