Anomaly Injection

Anomaly Injection

Great Generator creates clean relationship-safe data by default. Anomaly injection is separate and opt-in.

data = generate_domain(
    "ecommerce",
    anomalies={
        "null_rate": 0.02,
        "duplicate_rate": 0.01,
        "orphan_fk_rate": 0.001,
        "late_arrival_rate": 0.02,
        "outlier_rate": 0.005,
        "invalid_status_rate": 0.01,
    },
)

Supported anomaly types include:

• nulls
• duplicates
• invalid values
• outliers
• late-arriving records
• out-of-order records
• broken references / orphan foreign keys
• negative amounts
• skew

Keeping anomalies separate from realistic base generation matters. It lets users decide whether they want clean demo data or intentionally messy data for data-quality testing.

Labeled anomaly ground truth

Use return_labels=True when you want an answer key for every planted defect.

data = generate_domain(
    "ecommerce",
    anomalies={"null_rate": 0.02, "invalid_status_rate": 0.005},
    return_labels=True,
)

labels = data["_anomaly_labels"]

The label table includes:

• table
• row_index
• primary_key
• primary_key_value
• column
• anomaly_type
• original_value
• corrupted_value

This is useful for data-quality demos, anomaly detection benchmarks, and QA tests.