Skip to content

Ordinal Uncertain Ground Truth Comparison

Script: examples/ordinal_uncertain_ground_truth_comparison.py

→ API: OrdinalCrossEntropyLoss, CumulativeLinkLoss, PseudoLabelConsistencyLoss.

Compares regression-as-classification methods when ordered labels are ambiguous and represented as soft bin probabilities.

Methods:

  • HardOrdinalCE
  • SoftOrdinalCE
  • SoftOrdinalCE+Pseudo
  • SoftCumulativeLink

What it demonstrates:

  • plausibility-style soft ordered-bin targets on the labeled subset
  • confidence-gated soft pseudo labels from a teacher model
  • direct use of the existing ordinal losses with soft targets

Reported metrics:

  • Accuracy
  • OrdinalMAE
  • QWK
  • TrueNLL
  • PlausibilityCE
  • PseudoAcceptRate
  • train_s, eval_s

Run

pixi run python examples/ordinal_uncertain_ground_truth_comparison.py

Summary Artifact

pixi run python examples/ordinal_uncertain_ground_truth_comparison.py \
  --summary-json-path reports/example_summaries/ordinal_uncertain_ground_truth_comparison_full.json