Dataset onboarding and model-quality review

A buyer should know what data was evaluated, which engine produced the result, what the held-out metrics say, and what is still unknown before trusting the workflow.

Process a scan and get a free report

Dataset provenance

Dataset
AI4Shipwrecks
Status
Local run artifact available; external dataset provenance exists only as dataset_id in this metrics file.
Artifact
artifacts/metrics/ai4shipwrecks_sonar_metrics.json

Engine and split

Engine label
trained_model
Split
official_train_test
Split caveat
Held-out evaluation split reported by the artifact. Site-level leakage audit is not exposed to the UI yet.

Run artifact

Run ID
sonar_train_20260512T210538Z
Timestamp
5/12/2026, 9:05:38 PM UTC
Selected checkpoint
artifacts/models/sonar_unet_ai4shipwrecks_pw16.pt

Held-out metrics

MetricValueSource
IoU0.285artifacts/metrics/ai4shipwrecks_sonar_metrics.json
Dice/F10.347artifacts/metrics/ai4shipwrecks_sonar_metrics.json
Precision0.366artifacts/metrics/ai4shipwrecks_sonar_metrics.json
Recall0.699artifacts/metrics/ai4shipwrecks_sonar_metrics.json

Positive-label metrics

MetricValueSource
IoU0.218artifacts/metrics/ai4shipwrecks_sonar_metrics.json
Dice/F10.320artifacts/metrics/ai4shipwrecks_sonar_metrics.json
Precision0.349artifacts/metrics/ai4shipwrecks_sonar_metrics.json
Recall0.511artifacts/metrics/ai4shipwrecks_sonar_metrics.json

Label mix

74 of 120 samples contain positive labels (62%).

93 of 120 samples contain positive predictions (78%).

18 samples were empty in both prediction and label.

False-positive / false-negative caveats

  • Metrics are local run artifacts, not validated commercial performance.
  • False-positive and false-negative example lists are not present in this artifact, so the UI only shows aggregate caveats.
  • Dataset license, scene/site split proof, and label-review status are not exposed in the current frontend contract.
  • Positive-label metrics are shown separately because all-empty samples can make aggregate segmentation scores look cleaner than target-bearing data.

Next action for a messy dataset

  1. Bring one messy operator dataset with source, license/permission notes, sensor metadata, and any available labels.
  2. Run ingestion validation before analysis: file type, size, decode path, modality, and source ownership.
  3. Create a held-out split by site or mission pass before training or evaluation.
  4. Review false-positive and false-negative examples with the operator before presenting model quality as useful for planning.