Python SDK

Leakage Models

SDK Reference: leakage detection objects

Target leakage occurs when a feature carries information from the future or from the target itself, giving a model an unfair advantage at training time. The review.leakage reviewer runs named pattern detectors against the frozen profile; each detector emits LeakageFinding objects.

LeakageFinding

python

1@dataclass(frozen=True, slots=True)
2class LeakageFinding:
3    pattern: str                 # detector ID that produced this finding
4    column_name: str
5    title: str
6    rationale: str
7    evidence: Mapping[str, Any]
8    confidence: float            # 0.0 to 1.0
9    severity: str                # "info" | "warning" | "critical"
10    suggested_action: str

The Six Built-in Detectors

Pattern ID	Detector	What It Flags
target_correlation	Target Correlation	Columns with extreme correlation to the declared target.
identifier	Identifier Shape	ID-like columns that also correlate with the target.
timestamp	Timestamp Leakage	Datetime columns that extend past a declared prediction cutoff.
future_info	Future Information	Columns named like the outcome, or datetime columns extending past a declared event timestamp.
duplicate_target	Duplicate Target Information	Columns that are a near-deterministic copy or transform of the target.
suspicious_correlation	Suspicious Correlation	Suspicious correlations with a secondary signal, never magnitude alone.

Detectors run through builtin_detectors() and are supplied to the review.leakage reviewer. Detector findings carry a confidence label: High at or above 0.7, Medium at or above 0.4, and Low below that.

From Detectors to Review Findings

The leakage reviewer merges detector findings into the shared RuleFinding schema so every section speaks the same language:

One pattern on a column becomes a finding with rule_id = "leakage.<pattern>".
Several patterns on the same column are merged into a single finding with rule_id = "leakage.multiple_patterns", keeping the worst severity and the highest confidence.
The original pattern, confidence level, rationale, and suggested action are preserved in the finding's evidence and metadata.