Skip to content

Dataset Split and Leakage Controls by Scale

Last updated: 2026-05-24

Dataset splits are release artifacts, not convenience files. In autonomy, a train/validation/test split can change whether a model is genuinely learning generalizable perception or only memorizing routes, vehicles, map tiles, labeler habits, repeated clips, or future information. The risk increases when training data comes from aggregated LiDAR maps, LiDAR-image fusion, map-derived pseudo-labels, active learning, ML-related SLAM, synthetic data, and multi-site fleets.

This page defines split and leakage controls across MLOps scale levels. Use it when a dataset, label batch, replay package, feature snapshot, embedding index, or model release needs evidence that evaluation data stayed independent from training and tuning decisions.


Core Principle

The split manifest is the firewall between learning and evidence. It should be immutable for any promoted baseline, candidate, replay package, safety claim, or benchmark result.

The practical rule:

  1. Split before training-time fitting, feature selection, normalization, labeler calibration, threshold tuning, and synthetic augmentation decisions.
  2. Assign groups, not only frames, when frames share route, site, session, object identity, map tile, labeler, source map, or event family.
  3. Treat validation and test sets as consumers with allowed uses. A validation set may tune thresholds. A release test, local holdout, safety holdout, or benchmark set must not.
  4. Store split lineage next to the dataset snapshot, model registry record, replay package, and release packet.
  5. Rerun leakage checks whenever a source map, calibration, taxonomy, pseudo-label batch, feature materialization, embedding index, synthetic generator, or federated client cohort changes.

Scikit-learn's leakage guidance is a useful baseline: split first, do not fit preprocessing on test data, and use pipelines to keep transformations scoped to training data. Fleet autonomy needs the same discipline plus grouping and lineage for site, map, vehicle, sensor, route, labeler, and time dependencies.


Scope

ArtifactLeakage questionTypical split unit
Single-scan LiDAR segmentationDid adjacent frames, same object tracks, or same session leak across train/eval?Clip, session, route, vehicle, time window
LiDAR-image segmentationDid image-derived labels or calibration/projection artifacts leak into evaluation?Sensor kit, calibration package, projection batch, session
Aggregated-map segmentationDid map tiles, source-map sessions, or map-derived pseudo-labels appear in both training and release evidence?Map tile, source map, semantic layer, source session, site
Dynamic/static map cleaningDid removed dynamic residuals, stationary people, or static-transient assets appear in both training and hygiene evaluation?Track/event family, map tile, review batch
ML-related SLAMDid learned registration, place recognition, loop closure, or learned map priors train on evaluation routes or maps?Route, place cluster, map version, sensor kit, date
Replay and scenario miningDid replay scenarios become training examples after they were used as release gates?Scenario ID, event family, root-cause cluster
Feature and embedding storesWere feature definitions or vector indices fit with future/test data?Event time, materialization snapshot, corpus/index build
Foundation-model labelersDid prompt packs, evaluator thresholds, or retrieval corpora see benchmark answers?Prompt/evaluator version, corpus snapshot, labeler run
Federated or local trainingDid client/site holdouts leak through aggregation, shared adapters, or central distillation?Client, site, round, adapter, holdout cohort
Synthetic dataDid generated cases use test-set geometry, labels, seeds, or scene reconstructions?Generator seed, source asset, scenario template, domain randomization batch

Split Unit Taxonomy

Choosing the wrong split unit is the most common leakage source. A random frame split is rarely valid for autonomy release evidence because adjacent frames, repeated routes, static background, and object identities are highly correlated.

Split unitUse whenLeakage prevented
FrameOnly for S0 smoke checks on independent static samplesMinimal; not release grade for temporal logs
Clip/windowAdjacent frames share scene contextImmediate temporal duplicate leakage
Trip/sessionVehicle drives one continuous route or taskRoute context, lighting, operator behavior, repeated actors
Route/zoneManaged-site tasks repeat the same lanes, stands, aisles, gates, or quaysRoute memorization and local geometry shortcuts
Site/customerNew airport, port, campus, warehouse, or district is a deployment targetSite overfitting and customer-specific procedure leakage
Vehicle/sensor kitHardware, firmware, LiDAR model, camera layout, or mounting differsCalibration and sensor-bias leakage
Calibration packageProjection, fusion, or back-projection is calibration-dependentLiDAR-image or map-label projection leakage
Map version/tileLabels or priors come from registered mapsSource-map and semantic-layer leakage
Time/date/AIRAC cycleFuture conditions differ from training conditionsFuture leakage, seasonal leakage, construction or map-change leakage
Event family/root causeMultiple clips are mined from one incident or anomaly clusterReplay scenario duplication
Object or actor identityPersonnel, vehicles, aircraft, assets, or static objects repeatIdentity memorization
Labeler/prompt/evaluatorAuto-labelers, VLMs, LLM judges, or human reviewer pools differLabeler-style leakage and benchmark contamination
Synthetic source asset/seedSynthetic scenes are generated from real assets or templatesTemplate and seed leakage
Federated client/siteTraining happens across site or tenant boundariesClient holdout and privacy-evidence leakage

Leakage Mode Taxonomy

Leakage modeExampleControl
Temporal future leakageModel trains on data collected after the validation/test periodTime-ordered split, cutoff timestamp, exclusion window
Adjacent-frame leakageEvery fifth LiDAR frame goes to validation while neighboring frames trainClip/session grouping and temporal gap
Near-duplicate clip leakageSame hard-brake or FOD event is mined into both active learning and replayEvent-family grouping and scenario lineage
Route/site leakageModel evaluated on the same route geometry and signage it trained onRoute, zone, and site holdouts
Vehicle/sensor leakageEvaluation vehicle shares calibration quirks with training fleetHeld-out vehicles, sensor kits, and calibration packages
Map/tile leakageMap-derived labels from a source map train the model, then the same map tile evaluates itMap-version and tile-level split firewall
Source-map leakageLearned SLAM prior trains on the same source sessions used for geometry QASource-session and map-build lineage gates
Labeler leakageHuman or foundation-model labeler tuned on the benchmark answer setLabeler/prompt/evaluator registry and allowed-use state
Pseudo-label leakageBack-projected semantic-map labels enter training and release eval from the same source mapPseudo-label batch manifest and split checks
Active-learning leakageRelease failures are mined into training, then reused unchanged as release gatesReplay scenario retirement or separate clean holdout
Feature leakageFeature normalization, PCA, embedding index, or retrieval corpus is fit on eval/test examplesFit transforms on training-only snapshots; record corpus/index build
Monitoring leakageDelayed labels from production monitoring become both retraining data and post-release success evidenceMonitoring cohort assignment and delayed-label evidence separation
Synthetic leakageSynthetic generator reconstructs or copies test-set scenesSeed/source-asset split and real-only final validation
Federated leakageA held-out client contributes updates, adapters, or teacher logits to a global modelClient-holdout policy and round-level aggregation manifest
Benchmark leakagePublic leaderboard or internal safety test informs hyperparameters repeatedlyEvidence-locked test set, evaluation budget, and blind final gate

Scale Ladder

MLOps scaleMinimum split controlLeakage controlPromotion blocker
S0 notebook researchFixed train/validation/test file or seed in the run noteManual check that examples are not obvious duplicatesResult reused as a baseline without data pointer and split note
S1 repeatable prototypeImmutable dataset snapshot plus split manifestRepeatable split generator with clip/session groupingBaseline metric changes without frozen split ID
S2 single-product productionRelease dataset manifest with train/eval/replay split IDsAutomated duplicate, temporal, feature-fit, label, and replay overlap checksCandidate lacks clean split report tied to registry version
S3 fleet and multi-siteSite/route/vehicle/map local holdouts and ODD-cell split manifestsLocal holdout protection, event-family grouping, active-learning/replay separationGlobal aggregate passes while target site holdout is contaminated
S4 regulated safety-criticalEvidence-locked safety holdouts with retention and waiver policyTest-set access logging, evaluation budget, incident freeze, safety-case traceSafety claim uses data that influenced training, tuning, labels, or thresholds
S5 platform scaleShared split service and lineage graph across products and tenantsPolicy-as-code, split-as-contract API, cross-tenant isolation, leakage audit APIPlatform allows unsupported reuse or ambiguous split ownership

Small fleets can need S4 split controls when the model affects people, aircraft, protected zones, or compliance claims. Large offline research programs can remain S1 if outputs never cross into release authority.


Split Manifest Contract

At S1 the manifest can be a JSON sidecar. At S2-S5 it should become a cataloged artifact with lineage, owner, digest, policy result, and registry links.

FieldRequirement
split_idImmutable ID for the split assignment
taskDetection, semantic segmentation, map segmentation, SLAM registration, replay, labeler evaluation, or foundation-model judging
artifact_scopeDataset, label batch, replay pack, feature snapshot, embedding index, pseudo-label export, or model release
dataset_snapshot_idImmutable data/lakehouse/DVC/Iceberg snapshot
split_versionSplit generator version and parameters
assignment_unitFrame, clip, session, route, site, vehicle, map tile, client, event family, or combined group
grouping_keysSite, route, session, vehicle, sensor kit, calibration, map version, map tile, object identity, event family, prompt pack
temporal_cutoffTraining cutoff time and validation/test time windows
exclusion_windowsMinimum time gap around split boundaries and event families
holdout_scopeSites, routes, vehicles, map versions, clients, weather bins, or ODD cells reserved for validation/test/safety
source_map_scopeSource map manifest, semantic layer, map-hygiene layers, source sessions, and tile IDs for map-derived data
labeler_scopeHuman team, vendor batch, model labeler, prompt pack, evaluator, threshold set, retrieval corpus
feature_scopeFeature definition, materialization snapshot, event-time policy, embedding model, corpus, index build
synthetic_scopeGenerator version, source assets, seeds, scenario templates, domain randomization policy
privacy_and_residencyTenant/site restrictions, deletion state, allowed jurisdictions, retention class
allowed_useTrain, validation, tuning, release test, safety holdout, replay, benchmark, monitoring, review only
leakage_report_idOutput of duplicate, temporal, lineage, feature-fit, labeler, replay, and policy checks
downstream_consumersTraining runs, registry versions, release packets, safety-case claims, replay suites
owner_and_expiryData owner, model owner, approver, expiry/revalidation trigger

Control Gates

GateRequired checkBlocking condition
Training exportSplit manifest exists before examples are materializedDataset is drawn from mutable paths or split is generated after labels/features are fit
PreprocessingTransform state is fit only on training partitionNormalizer, imputer, PCA, class weights, or feature selector uses eval/test data
Label importLabel batches preserve allowed-use and split stateCandidate/review labels become release truth or test answers leak to labelers
Pseudo-label promotionSource-map, taxonomy, calibration, release-state, and split IDs are stableMap-derived labels share source tiles with release evaluation without holdout policy
Feature/embedding buildEvent-time join and corpus/index lineage are recordedFuture/test examples influence feature materialization or retrieval corpus
Replay promotionScenario IDs and event families are not already training examples unless marked retiredRelease replay suite overlaps training or active-learning examples
Federated aggregationClient/site holdouts are excluded from updates, distillation, and shared adaptersHeld-out clients contribute model updates or teacher logits
Release evaluationCandidate registry record names split IDs and leakage reportTest or local holdout set was used for tuning or repeated threshold selection
Safety evidenceHoldout is evidence-locked with access log, owner, expiry, and retentionSafety claim depends on data that influenced training, map labels, or evaluation policy
Platform reuseSplit ownership and allowed use are queryable by APIAnother team reuses a release test, tenant holdout, or benchmark set as training data

Training and Evaluation Architecture Comparison

ArchitectureAdvantagesDisadvantagesBest use
Random frame splitSimple, fast, useful for pipeline smoke testsSevere leakage for video/LiDAR sequences, static maps, and repeated sitesS0 only, non-release debugging
Temporal splitMirrors production by training on past data and testing on later dataDoes not prove route/site/vehicle generalization by itselfS1-S2 baselines, drift studies, monitoring retrain decisions
Temporal split with gapReduces adjacent-frame and same-object leakageRequires enough data volume; gap choice must be justifiedLiDAR/image logs, airside routes, event-triggered clips
Group split by session/routePrevents route and repeated-scene memorizationCan produce imbalanced class/site distributionsManaged-site perception and replay packages
Vehicle/sensor-kit holdoutTests hardware and calibration robustnessExpensive for small fleets; confounds vehicle and site if fleet is not balancedS2+ release for heterogeneous fleets
Site/ODD-cell holdoutStrong evidence for new airport, port, campus, warehouse, or district transferHardest split; may make metrics lower and data needs higherS3+ multi-site release and local acceptance
Map-version/tile holdoutTests generalization to unseen map geometry and source-map buildsNeeds careful tile boundary handling to avoid seam leakageAggregated-map segmentation and ML-SLAM QA
Event-family holdoutPrevents one incident from training and validating the same root causeRequires event clustering and scenario lineageSafety replay, active learning, root-cause regression suites
Group k-fold / leave-one-group-outUses limited data efficiently while respecting group independenceMore compute; fold leakage if grouping keys are incompleteS1-S3 model comparison and ablation
Evidence-locked blind holdoutHighest credibility for release or safety claimsSlow, expensive, and access-controlled; unsuitable for daily tuningS4 safety evidence and benchmark-like final gates
Synthetic-real splitTests whether synthetic data helps real data without replacing itSynthetic assets can leak real test geometry if poorly sourcedRare hazards, weather, FOD, digital-twin transfer
Federated client holdoutTests cross-client/site generalization and privacy-preserving trainingHard to debug; aggregation can indirectly contaminate holdoutsS3-S5 local, hybrid, federated, or privacy-constrained training

For scikit-learn-scale experiments, GroupKFold, LeaveOneGroupOut, and TimeSeriesSplit are useful mental models. At autonomy scale, the same ideas must be lifted into cataloged manifests so every pipeline and release gate sees the same split policy.


LiDAR, Image, and ML-SLAM Specific Rules

LiDAR

  • Do not split individual point frames randomly when frames come from dense multi-sweep or aggregated-map sessions.
  • Group by clip/session/route and keep a temporal gap for adjacent scans.
  • Record LiDAR model, firmware, intensity calibration, mounting position, and extrinsic calibration in the split manifest.
  • Treat de-skewing, ground removal, dynamic-object removal, voxel statistics, and range normalization as training-fitted or source-dependent operations that require lineage.

Image and LiDAR-Image Fusion

  • Record camera model, intrinsics, extrinsics, time-sync package, projection code, and image source batch.
  • Keep image-dependent labels separate from LiDAR-only release evidence unless the runtime artifact also depends on images.
  • If images are used only for train-time distillation or colorized map labeling, record that as a training-only dependency with explicit allowed use.
  • Rebuild or quarantine split evidence when projection QA, calibration, or time-sync changes.

Aggregated Maps and Semantic Maps

  • Split by source map, map version, semantic layer, tile, source session, and release-state mask.
  • Do not let a semantic map tile produce training labels and also serve as the independent release test for that same product mode.
  • Preserve map-hygiene labels for permanent_static, dynamic_residual, movable_static, static_transient, fod_candidate, artifact, and unknown_review so release-state leakage is measurable.
  • Rerun leakage checks when map merging, loop closure, source-map acceptance, dynamic residual removal, or static-transient quarantine changes.
  • Learned place recognition and registration models should hold out place clusters, routes, and map versions, not only individual pairs.
  • Learned dynamic removal and map cleaning should evaluate on scenes whose dynamic/static decisions were not used to tune thresholds.
  • Neural or Gaussian map priors should record source sessions and reconstruction assets so evaluation maps are not regenerated from training scenes.
  • SLAM improvements that change poses or map geometry invalidate downstream map-derived split claims until affected labels and replay packages are rechecked.

Managed-Site and Non-Road Urban District Rules

Airside, port, yard, campus, warehouse, terminal-frontage, facade, utility, and other non-road mapping programs have repeated geometry and strong local procedures. That makes leakage easy and false confidence common.

Domain featureSplit implication
Repeated routes and standsHold out routes/zones, not random frames
Stationary people or parked movable assetsUse release-state labels so static-but-transient objects do not become permanent training truth
Construction, de-icing, temporary closuresSplit by date/map version and expire old holdouts after operational change
Thin utility/facade infrastructureUse site and asset-family holdouts to avoid memorizing repeated structures
Local procedure differencesRelease evidence must be scoped by ODD cell and operating phase
Multi-tenant sitesSplit and access policy must preserve tenant/customer boundaries
Public-proxy benchmarksTreat as pretraining or method evidence, not local release evidence

The managed-site default should be: global training pool for representation learning, local holdouts for release evidence, and evidence-locked safety holdouts for behavior authority.


Scorecards

KPIMeaningRelease interpretation
Split manifest coveragePercentage of training/eval/replay artifacts with split IDsS1 baseline and S2 release blocker
Group-key completenessRequired grouping keys present for site, route, session, vehicle, map, labeler, and event familyMissing key means leakage check is incomplete
Temporal gap violationsAdjacent train/eval examples within exclusion windowBlock release eval until removed or waived
Duplicate/event overlap rateNear-duplicate or same-event examples across train/eval/replayBlock if release replay or local holdout is affected
Feature-fit leakage ratePreprocessing or feature materialization fit on eval/testBlock candidate; rebuild transform
Map/tile overlap rateShared source-map tiles or sessions across train and independent evalBlock map-derived release evidence
Labeler contamination rateBenchmark/eval answers exposed to labelers, prompts, evaluators, or reviewersBlock affected labels/evals
Holdout access countNumber of human or automated accesses to blind/evidence holdoutS4/S5 audit input; high use may retire the holdout
Retired-test reuse countEvidence tests moved into training after retirement and replacementHealthy if explicit; dangerous if silent
Waiver ageAge of unresolved split/leakage exceptionOld waivers should block S4/S5 release

Failure Modes

Failure modeConsequenceControl
Split generated after preprocessingTest information changes model featuresSplit first; fit transforms on train only
Random frames from one route split across train/testMetrics overstate route generalizationClip, session, route, and temporal grouping
Local holdout enters active-learning trainingSite release evidence becomes self-fulfillingHoldout allowed-use policy and lineage block
Replay scenario used for training without retirementCandidate learns the release gateScenario lineage and replacement replay case
Map-derived labels from eval tile train the modelAggregated-map metric measures memorizationMap/tile/source-session split firewall
Static-transient object treated as permanent ground truthModel learns stationary people or parked assets as map fixturesMap-hygiene release-state masks and static-transient quarantine
Feature/vector index includes future examplesRetrieval or feature quality leaks test dataEvent-time materialization and corpus/index split policy
Prompt/evaluator sees test answersFoundation-model labeler or judge contaminates benchmarkPrompt/evaluator registry and blind eval discipline
Federated held-out client contributes updatesClient holdout no longer independentClient-round manifest and aggregation policy
Old safety holdout reused too oftenSafety evidence becomes tuned to the known test setAccess log, evaluation budget, holdout retirement
Split policy differs across teamsModel, data, replay, and safety evidence cannot be reconciledShared split service or cataloged split-as-contract

  • mlops-scale-research-scope.md - MLOps maturity ladder and research backlog.
  • mlops-scorecards-and-kpis-by-scale.md - release-blocking metrics and operating cadence.
  • mlops-reference-architectures-by-scale.md - S0-S5 architecture patterns and platform interfaces.
  • mlops-migration-checklist-by-scale.md - transition evidence before raising artifact authority.
  • site-sliced-release-evidence-by-scale.md - ODD-cell manifests and local holdout evidence.
  • model-governance-release-evidence.md - release packet, registry aliases, and rollback evidence.
  • map-derived-pseudo-label-invalidation-protocol.md - semantic-map pseudo-label state machine and impact graph.
  • feature-embedding-store-ops-by-scale.md - point-in-time feature and vector-index leakage controls.
  • federated-privacy-preserving-training-policy-by-scale.md - client/site holdout rules for federated and hybrid training.
  • ../data-platform/fleet-data-pipeline.md - raw logs, split generation, and training export pipeline.
  • ../data-platform/data-catalog-lineage-quality-ops.md - catalog states, lineage, and data product promotion.
  • ../data-platform/perception-slam-fleet-data-contract.md - perception/SLAM data contracts and duplicate grouping.
  • ../../10-knowledge-base/machine-learning/av-data-evaluation-fundamentals.md - AV dataset and evaluation fundamentals.
  • ../../10-knowledge-base/machine-learning/evaluation-calibration-and-data-leakage-first-principles.md - first principles for leakage and calibration.
  • ../../30-autonomy-stack/perception/overview/aggregated-map-semantic-segmentation.md - semantic-map training exports and release-state controls.
  • ../../30-autonomy-stack/localization-mapping/overview/ml-related-slam-research-scope.md - learned SLAM and map-cleaning research scope.

Sources

Public research notes collected from public sources.