# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

"""Accuracy per question type, per category, and overall, from eval/run_eval.py output.

Accuracy is correct / (correct + wrong). Replies that named no single choice
(parse_failure) and requests that failed (error) are counted and reported but
left out of the accuracy.

Usage:
    python eval/score.py results/<model>
"""

from __future__ import annotations

import argparse
import json
from collections import Counter, defaultdict
from pathlib import Path

# Type value in the data -> (name used in the paper, category)
TYPES = {
    "audio_presence": ("Sound Presence", "Audio"),
    "audio_only_recall": ("Audio Scene Recall", "Audio"),
    "audio_occasion": ("Audio Occasion", "Audio"),
    "visual_detail_recall": ("Visual Detail Recall", "Visual"),
    "preference_state_at_time": ("Past Preference Recall", "Personalization"),
    "preference_change_count": ("Preference Change Count", "Personalization"),
    "preference_replaced": ("Preference Replacement", "Personalization"),
    "preference_never_asked": ("Preference Hallucination", "Personalization"),
    "exception_recall": ("Exception Recall", "Personalization"),
    "exception_expiry": ("Exception Expiry", "Personalization"),
    "hearing_adjustment": ("Adjustment Suggestion", "Hearing Adjustment"),
}
CATEGORIES = ["Audio", "Visual", "Personalization", "Hearing Adjustment"]


def accuracy(counts: Counter) -> str:
    scored = counts["correct"] + counts["wrong"]
    return f"{100 * counts['correct'] / scored:6.2f}" if scored else "     -"


def main() -> None:
    parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
    parser.add_argument("results_dir", type=Path)
    args = parser.parse_args()

    by_type: dict[str, Counter] = defaultdict(Counter)
    by_category: dict[str, Counter] = defaultdict(Counter)
    overall: Counter = Counter()
    for path in sorted(args.results_dir.glob("*.jsonl")):
        latest: dict[str, dict] = {}
        for line in path.read_text().splitlines():
            record = json.loads(line)
            latest[record["candidate_id"]] = record  # a resumed run's retry replaces the error
        for record in latest.values():
            name, category = TYPES[record["type"]]
            by_type[name][record["outcome"]] += 1
            by_category[category][record["outcome"]] += 1
            overall[record["outcome"]] += 1

    print(f"{'type':28s} {'acc':>6s}  {'n':>5s}")
    for key, (name, _) in TYPES.items():
        if name in by_type:
            print(f"{name:28s} {accuracy(by_type[name])}  {sum(by_type[name].values()):5d}")
    print()
    for category in CATEGORIES:
        if category in by_category:
            print(f"{category:28s} {accuracy(by_category[category])}  {sum(by_category[category].values()):5d}")
    print(f"{'Overall':28s} {accuracy(overall)}  {sum(overall.values()):5d}")
    print(f"\nparse failures: {overall['parse_failure']}, errors: {overall['error']}")


if __name__ == "__main__":
    main()