scorer.py

import pandas as pd
from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score, confusion_matrix


def _resolve_target_col(df: pd.DataFrame) -> str:
    label_cols = [c for c in df.columns if c.startswith("label_")]
    if len(label_cols) == 1:
        return label_cols[0]

    raise ValueError(
        f"Expected exactly one label column starting with 'label_', found: {label_cols}"
    )


def score(solution, submission):
    if isinstance(solution, str):
        solution = pd.read_csv(solution)

    if isinstance(submission, str):
        submission = pd.read_csv(submission)

    target_col = _resolve_target_col(solution)

    if target_col not in submission.columns:
        submission_label_cols = [c for c in submission.columns if c.startswith("label_")]
        if len(submission_label_cols) == 1:
            submission = submission.rename(columns={submission_label_cols[0]: target_col})
        else:
            raise ValueError(
                f"Submission must contain target column '{target_col}' or exactly one label column. "
                f"Found: {list(submission.columns)}"
            )

    y_true = solution[target_col]
    y_pred = submission[target_col]

    accuracy = accuracy_score(y_true, y_pred)
    precision = precision_score(y_true, y_pred, zero_division=0)
    recall = recall_score(y_true, y_pred, zero_division=0)
    f1 = f1_score(y_true, y_pred, zero_division=0)
    cm = confusion_matrix(y_true, y_pred)

    return {
        "target_column": target_col,
        "positive_class": 1,
        "primary_metric": "recall_cascade_detection",
        "accuracy": accuracy,
        "precision": precision,
        "recall_cascade_detection": recall,
        "false_safe_rate": 1 - recall,
        "f1": f1,
        "confusion_matrix": cm.tolist(),
    }