MLVersionComparison
An earlier version's score on a new run's held-out rows
Fields
The earlier version that was re-scored; null when an experiment run was
The earlier ML experiment run that was re-scored; null when a version was
Short display label of what was re-scored (e.g. v3, or run 3 for an experiment run)
Whether this score can be read against the new one
Possible values (4)
COMPARABLEThe scores can be read against each other
PIPELINE_CHANGEDScored, but the earlier version read its rows through a different data pipeline
OVERLAP_RISKScored, but some of the earlier version's training rows may be in this holdout, which would flatter its score
INCOMPATIBLENot scored: the earlier version predicts something else or cannot read these rows
Code explaining a status other than COMPARABLE: TARGET_CHANGED, MODEL_TYPE_CHANGED, MISSING_COLUMNS, ARTIFACT_MISSING, SCORING_FAILED, PIPELINE_CHANGED or SPLIT_OVERLAP
The earlier version's scores on this run's held-out rows; null when it was not scored
Referenced by (1)
Exact type SDL
"""An earlier version's score on a new run's held-out rows"""
type MLVersionComparison {
"""
The earlier version that was re-scored; null when an experiment run was
"""
versionId: MLModelConfigVersionID
"""
The earlier ML experiment run that was re-scored; null when a version was
"""
runId: MLExperimentRunID
"""
Short display label of what was re-scored (e.g. v3, or run 3 for an experiment run)
"""
label: String!
"""Whether this score can be read against the new one"""
status: MLComparability!
"""
Code explaining a status other than COMPARABLE: TARGET_CHANGED, MODEL_TYPE_CHANGED, MISSING_COLUMNS, ARTIFACT_MISSING, SCORING_FAILED, PIPELINE_CHANGED or SPLIT_OVERLAP
"""
reason: String
"""
The earlier version's scores on this run's held-out rows; null when it was not scored
"""
metrics: TrainingMetrics
}