evaluate.py

#!/usr/bin/env python3
"""Official deterministic evaluator for SABRE-Prior."""

from __future__ import annotations

import argparse
import json
import re
from collections import Counter, defaultdict
from pathlib import Path
from typing import Any


SUBSETS = ("context", "texture", "attribute", "language")
PAIRED_PROBES = {
    "context": ("base_source", "base_target", "edited_source", "edited_target"),
    "texture": (
        "base_normal",
        "base_counterfactual",
        "edited_normal",
        "edited_counterfactual",
    ),
}
NUMBER_WORDS = {
    "zero": "0",
    "one": "1",
    "two": "2",
    "three": "3",
    "four": "4",
    "five": "5",
    "six": "6",
    "seven": "7",
    "eight": "8",
    "nine": "9",
    "ten": "10",
    "eleven": "11",
    "twelve": "12",
    "thirteen": "13",
    "fourteen": "14",
    "fifteen": "15",
    "sixteen": "16",
    "seventeen": "17",
    "eighteen": "18",
    "nineteen": "19",
    "twenty": "20",
}


def load_jsonl(path: Path) -> list[dict[str, Any]]:
    if not path.is_file():
        raise FileNotFoundError(path)
    rows: list[dict[str, Any]] = []
    for line_number, line in enumerate(path.read_text(encoding="utf-8").splitlines(), 1):
        if not line.strip():
            continue
        try:
            row = json.loads(line)
        except json.JSONDecodeError as exc:
            raise ValueError(f"Invalid JSON in {path}:{line_number}: {exc}") from exc
        if not isinstance(row, dict):
            raise ValueError(f"Expected an object in {path}:{line_number}")
        rows.append(row)
    return rows


def index_unique(rows: list[dict[str, Any]], path: Path) -> dict[str, dict[str, Any]]:
    indexed: dict[str, dict[str, Any]] = {}
    for row in rows:
        item_id = str(row.get("id") or "").strip()
        if not item_id:
            raise ValueError(f"A row in {path} has no id")
        if item_id in indexed:
            raise ValueError(f"Duplicate id in {path}: {item_id}")
        indexed[item_id] = row
    return indexed


def normalize_yes_no(value: Any) -> str:
    text = str(value or "").strip().casefold()
    if text.startswith("yes"):
        return "yes"
    if text.startswith("no"):
        return "no"
    return "unknown"


def normalize_count(value: Any) -> str:
    text = str(value or "").strip().lower().replace("×", "x")
    text = re.sub(r"[^a-z0-9+\- x]+", " ", text)
    text = re.sub(r"\s+", " ", text).strip()
    text = text.split(" instead of ", 1)[0].strip()

    pair = re.search(r"\b(\d+)\s+and\s+(\d+)\b", text)
    if pair:
        return f"{pair.group(1)} and {pair.group(2)}"
    for word1, digit1 in NUMBER_WORDS.items():
        for word2, digit2 in NUMBER_WORDS.items():
            if re.search(rf"\b{word1}\s+and\s+{word2}\b", text):
                return f"{digit1} and {digit2}"

    grid = re.search(r"\b(\d+)\s*(?:x|by|-by-)\s*(\d+)\b", text)
    if grid:
        return f"{grid.group(1)}x{grid.group(2)}"
    for word1, digit1 in NUMBER_WORDS.items():
        for word2, digit2 in NUMBER_WORDS.items():
            if re.search(rf"\b{word1}\s*(?:x|by|-by-)\s*{word2}\b", text):
                return f"{digit1}x{digit2}"

    numbers = re.findall(r"\b\d+\b", text)
    if numbers:
        return numbers[0]
    for word, digit in NUMBER_WORDS.items():
        if re.search(rf"\b{word}\b", text):
            return digit
    return text


def normalize_choice(value: Any) -> str:
    match = re.search(r"(?:^|[^A-Z])([A-D])(?:[^A-Z]|$)", str(value or "").strip().upper())
    return match.group(1) if match else "UNKNOWN"


def ratio(correct: int, total: int) -> dict[str, Any]:
    accuracy = correct / total if total else None
    return {
        "correct": correct,
        "total": total,
        "accuracy": accuracy,
        "accuracy_percent": round(accuracy * 100, 1) if accuracy is not None else None,
    }


def checked_predictions(
    questions: dict[str, dict[str, Any]], path: Path
) -> dict[str, dict[str, Any]]:
    predictions = index_unique(load_jsonl(path), path)
    missing = sorted(set(questions) - set(predictions))
    extra = sorted(set(predictions) - set(questions))
    if missing:
        raise ValueError(f"Missing {len(missing)} predictions; first missing id: {missing[0]}")
    if extra:
        raise ValueError(f"Found {len(extra)} unknown predictions; first unknown id: {extra[0]}")
    for item_id, row in predictions.items():
        if "prediction" not in row and "raw_prediction" not in row:
            raise ValueError(f"Prediction row {item_id} has no prediction field")
    return predictions


def prediction_value(row: dict[str, Any]) -> Any:
    return row["prediction"] if "prediction" in row else row.get("raw_prediction", "")


def score_paired(
    subset: str,
    questions: dict[str, dict[str, Any]],
    predictions: dict[str, dict[str, Any]],
) -> dict[str, Any]:
    required_probes = PAIRED_PROBES[subset]
    by_pair: dict[str, dict[str, bool]] = defaultdict(dict)
    probe_correct: Counter[str] = Counter()

    for item_id, question in questions.items():
        pair_id = str(question.get("pair_id") or "")
        probe = str(question.get("probe") or "")
        if not pair_id or probe not in required_probes:
            raise ValueError(f"Invalid pair_id/probe for {item_id}")
        if probe in by_pair[pair_id]:
            raise ValueError(f"Duplicate probe {probe} in pair {pair_id}")
        prediction = normalize_yes_no(prediction_value(predictions[item_id]))
        expected = normalize_yes_no(question["answer"])
        correct = prediction == expected
        by_pair[pair_id][probe] = correct
        probe_correct[probe] += int(correct)

    for pair_id, probes in by_pair.items():
        if set(probes) != set(required_probes):
            raise ValueError(f"Pair {pair_id} does not contain exactly the four required probes")

    pair_correct = sum(all(probes[probe] for probe in required_probes) for probes in by_pair.values())
    question_correct = sum(probe_correct.values())
    return {
        "primary_metric": "strict_pair_accuracy",
        "primary": ratio(pair_correct, len(by_pair)),
        "diagnostics": {
            "question_accuracy": ratio(question_correct, len(questions)),
            "probe_accuracy": {
                probe: ratio(probe_correct[probe], len(by_pair)) for probe in required_probes
            },
        },
    }


def score_attribute(
    questions: dict[str, dict[str, Any]], predictions: dict[str, dict[str, Any]]
) -> dict[str, Any]:
    correct = sum(
        normalize_count(prediction_value(predictions[item_id]))
        == normalize_count(question["answer"])
        for item_id, question in questions.items()
    )
    return {"primary_metric": "exact_count_accuracy", "primary": ratio(correct, len(questions))}


def score_language(
    questions: dict[str, dict[str, Any]], predictions: dict[str, dict[str, Any]]
) -> dict[str, Any]:
    correct = sum(
        normalize_choice(prediction_value(predictions[item_id]))
        == normalize_choice(question["answer"])
        for item_id, question in questions.items()
    )
    return {"primary_metric": "exact_choice_accuracy", "primary": ratio(correct, len(questions))}


def evaluate_subset(root: Path, predictions_root: Path, subset: str) -> dict[str, Any]:
    metadata_path = root / "data" / subset / "metadata.jsonl"
    prediction_path = predictions_root / f"{subset}.jsonl"
    questions = index_unique(load_jsonl(metadata_path), metadata_path)
    predictions = checked_predictions(questions, prediction_path)
    if subset in PAIRED_PROBES:
        return score_paired(subset, questions, predictions)
    if subset == "attribute":
        return score_attribute(questions, predictions)
    if subset == "language":
        return score_language(questions, predictions)
    raise ValueError(f"Unknown subset: {subset}")


def main() -> int:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument(
        "--predictions",
        type=Path,
        required=True,
        help="Directory containing context.jsonl, texture.jsonl, attribute.jsonl, and language.jsonl.",
    )
    parser.add_argument("--dataset-root", type=Path, default=Path(__file__).resolve().parent)
    parser.add_argument("--subset", choices=("all", *SUBSETS), default="all")
    parser.add_argument("--output", type=Path)
    args = parser.parse_args()

    selected = SUBSETS if args.subset == "all" else (args.subset,)
    metrics: dict[str, Any] = {"metric_version": "sabre-prior-v1.0", "subsets": {}}
    for subset in selected:
        metrics["subsets"][subset] = evaluate_subset(
            args.dataset_root.resolve(), args.predictions.resolve(), subset
        )

    if args.subset == "all":
        accuracies = [metrics["subsets"][name]["primary"]["accuracy"] for name in SUBSETS]
        macro = sum(accuracies) / len(accuracies)
        metrics["macro_accuracy"] = macro
        metrics["macro_accuracy_percent"] = round(macro * 100, 1)

    output = json.dumps(metrics, indent=2, ensure_ascii=False) + "\n"
    if args.output:
        args.output.parent.mkdir(parents=True, exist_ok=True)
        args.output.write_text(output, encoding="utf-8")
    print(output, end="")
    return 0


if __name__ == "__main__":
    raise SystemExit(main())