tests.py

test_cases = [
    {
        "references": ["1111", "1111", "1111", "1111"],
        "predictions": ["1111", "1112", "1120", "1211"],
        "result": {
            "accuracy": 0.25,
            "hierarchical_precision": 1.0,
            "hierarchical_recall": 0.4,
            "hierarchical_fmeasure": 0.5714285714285715,
        },
    },
    {
        "references": ["1111"],
        "predictions": ["1111"],
        "result": {
            "accuracy": 1.0,
            "hierarchical_precision": 1.0,
            "hierarchical_recall": 1.0,
            "hierarchical_fmeasure": 1.0,
        },
    },
    {
        "references": ["1111"],
        "predictions": ["1112"],
        "result": {
            "accuracy": 0.0,
            "hierarchical_precision": 0.75,
            "hierarchical_recall": 0.75,
            "hierarchical_fmeasure": 0.75,
        },
    },
    {
        "references": ["1111"],
        "predictions": ["1120"],
        "result": {
            "accuracy": 0.0,
            "hierarchical_precision": 0.5,
            "hierarchical_recall": 0.5,
            "hierarchical_fmeasure": 0.5,
        },
    },
    {
        "references": ["1111"],
        "predictions": ["1211"],
        "result": {
            "accuracy": 0.0,
            "hierarchical_precision": 0.25,
            "hierarchical_recall": 0.25,
            "hierarchical_fmeasure": 0.25,
        },
    },
]