benchpress_eval/metrics.py

from __future__ import annotations

from typing import Any

import numpy as np
import pandas as pd

from .data import expand_triplets


def auroc_safe(y_true: pd.Series | np.ndarray, y_score: pd.Series | np.ndarray) -> float:
    """Rank-based AUROC with average ranks for ties."""
    y = pd.Series(y_true).astype(float)
    s = pd.to_numeric(pd.Series(y_score), errors="coerce")
    mask = np.isfinite(y) & np.isfinite(s)
    y = y[mask].astype(int)
    s = s[mask].astype(float)
    if y.nunique() < 2:
        return float("nan")
    n_pos = int((y == 1).sum())
    n_neg = int((y == 0).sum())
    if n_pos == 0 or n_neg == 0:
        return float("nan")
    ranks = s.rank(method="average")
    pos_rank_sum = float(ranks[y == 1].sum())
    return (pos_rank_sum - n_pos * (n_pos + 1) / 2.0) / (n_pos * n_neg)


def _score_column(predictions: pd.DataFrame, requested: str | None) -> str:
    if requested and requested in predictions.columns:
        return requested
    for candidate in ("score", "score_norm", "raw_score"):
        if candidate in predictions.columns:
            return candidate
    raise ValueError("Predictions must contain a score column, e.g. score or score_norm.")


def _standardize_negative_index(df: pd.DataFrame) -> pd.DataFrame:
    df = df.copy()
    if "negative_index" in df.columns:
        df["negative_index"] = pd.to_numeric(df["negative_index"], errors="coerce").fillna(-1).astype(int)
    return df


def align_predictions(
    dataset_df: pd.DataFrame,
    predictions: pd.DataFrame,
    score_col: str | None = None,
) -> pd.DataFrame:
    """Align model scores to EditJudge-Bench triplets.

    Supported prediction schemas:
    1. Expanded rows with label/edit_type/negative_type/score already present.
    2. Rows keyed by sample_id or parquet_row_index plus example_type and negative_index.
    3. Rows keyed by sample_id or parquet_row_index plus instruction text.
    """
    predictions = _standardize_negative_index(predictions)
    score_col = _score_column(predictions, score_col)
    predictions = predictions.rename(columns={score_col: "score"}).copy()
    predictions["score"] = pd.to_numeric(predictions["score"], errors="coerce")

    expanded_cols = {"label", "edit_type", "negative_type", "score"}
    if expanded_cols.issubset(predictions.columns):
        out = predictions.copy()
        if "ground_truth" not in out.columns:
            out["ground_truth"] = out["label"].astype(bool)
        return out

    triplets = expand_triplets(dataset_df)
    key = "sample_id" if "sample_id" in predictions.columns else "parquet_row_index"
    if key not in predictions.columns:
        raise ValueError("Predictions must contain sample_id or parquet_row_index.")

    if {"example_type", "negative_index"}.issubset(predictions.columns):
        join_cols = [key, "example_type", "negative_index"]
    elif "instruction" in predictions.columns:
        join_cols = [key, "instruction"]
    else:
        raise ValueError(
            "Predictions must include either example_type+negative_index or instruction."
        )

    keep_cols = join_cols + ["score"] + [
        col for col in ("raw_score", "raw_output", "model", "method") if col in predictions.columns
    ]
    merged = triplets.merge(predictions[keep_cols], on=join_cols, how="left", validate="one_to_one")
    missing = int(merged["score"].isna().sum())
    if missing:
        raise ValueError(f"Could not align {missing} triplets to prediction scores.")
    return merged


def compute_overall_metrics(aligned: pd.DataFrame) -> pd.DataFrame:
    per_edit = per_edit_type_auc(aligned)
    row: dict[str, Any] = {
        "global_auc": auroc_safe(aligned["label"], aligned["score"]),
        "macro_edit_auc": per_edit["auc"].mean(),
        "n_triplets": int(len(aligned)),
        "n_edits": int(aligned["sample_id"].nunique()) if "sample_id" in aligned.columns else np.nan,
    }
    return pd.DataFrame([row])


def per_edit_type_auc(aligned: pd.DataFrame) -> pd.DataFrame:
    rows = []
    for edit_type, group in aligned.groupby("edit_type", sort=True):
        rows.append(
            {
                "edit_type": edit_type,
                "auc": auroc_safe(group["label"], group["score"]),
                "n_triplets": int(len(group)),
                "n_edits": int(group["sample_id"].nunique()) if "sample_id" in group.columns else np.nan,
            }
        )
    return pd.DataFrame(rows)


def per_negative_type_auc(aligned: pd.DataFrame) -> pd.DataFrame:
    positives = aligned[aligned["label"] == 1]
    negatives = aligned[aligned["label"] == 0]
    rows = []
    for negative_type, neg_group in negatives.groupby("negative_type", sort=True):
        group = pd.concat([positives, neg_group], ignore_index=True)
        rows.append(
            {
                "negative_type": negative_type,
                "auc": auroc_safe(group["label"], group["score"]),
                "n_negative_triplets": int(len(neg_group)),
                "n_positive_triplets": int(len(positives)),
            }
        )
    return pd.DataFrame(rows)


def semantic_vs_noedit_summary(per_negative: pd.DataFrame) -> pd.DataFrame:
    semantic_names = {"counterfactual", "cross_type", "wrong_object"}
    semantic = per_negative[per_negative["negative_type"].isin(semantic_names)]
    no_edit = per_negative[per_negative["negative_type"].eq("no_edit")]
    return pd.DataFrame(
        [
            {
                "semantic_auc": semantic["auc"].mean(),
                "no_edit_auc": no_edit["auc"].mean(),
                "semantic_negative_types": ",".join(sorted(semantic["negative_type"].unique())),
            }
        ]
    )