"""Compare the two validation splits in Google's TyDi QA dataset on Hugging Face."""

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

import pyarrow.parquet as pq


INPUT_HASHES = {
    "primary": "2e9557a6ea4bfd65f381c64b91122c90554e4b00f86e15a8b4f1353be1c81510",
    "gold": "38760fbfc8fa4d0765026d56cfbb65a39231ad8cf5afef40f9c962a26bd3f94d",
}


def read_rows(path, split, columns):
    with path.open("rb") as source:
        if hashlib.file_digest(source, "sha256").hexdigest() != INPUT_HASHES[split]:
            raise ValueError(f"Unexpected {split} file digest: {path}")
    for batch in pq.ParquetFile(path).iter_batches(columns=columns, batch_size=256):
        yield from batch.to_pylist()


def label_type(annotations):
    if any(start >= 0 for start in annotations["minimal_answers_start_byte"]):
        return "has_text_answer_label"
    if any(label in {"YES", "NO"} for label in annotations["yes_no_answer"]):
        return "has_yes_no_label_only"
    if any(index >= 0 for index in annotations["passage_answer_candidate_index"]):
        return "has_passage_label_only"
    return "no_answer_label"


def compare(primary_path, gold_path):
    counts = defaultdict(Counter)
    primary_keys = defaultdict(list)
    for row in read_rows(primary_path, "primary", (
        "language", "document_title", "question_text", "annotations",
    )):
        language = row["language"]
        label = label_type(row["annotations"])
        primary_keys[language, row["document_title"], row["question_text"]].append(label)
        counts[language]["primary_rows"] += 1
        counts[language][f"primary_{label}"] += 1

    matched_keys = set()
    gold_ids = set()
    gold_question_keys = set()
    for row in read_rows(gold_path, "gold", ("id", "title", "question")):
        language, _, _ = row["id"].split("-", 2)
        if row["id"] in gold_ids:
            raise ValueError(f"Duplicate Gold Passage row ID: {row['id']}")
        gold_ids.add(row["id"])
        key = language, row["title"], row["question"]
        gold_question_keys.add(key)
        counts[language]["gold_rows"] += 1
        primary_matches = primary_keys.get(key, [])
        if not primary_matches:
            counts[language]["gold_rows_with_no_exact_match"] += 1
        elif len(primary_matches) > 1:
            counts[language]["gold_rows_with_ambiguous_match"] += 1
        else:
            counts[language]["gold_rows_with_unique_match"] += 1
            matched_keys.add(key)

    for language, _, _ in gold_question_keys:
        counts[language]["distinct_gold_question_keys"] += 1
    for key in matched_keys:
        language, _, _ = key
        counts[language]["matched_primary_rows"] += 1
        counts[language][f"matched_{primary_keys[key][0]}"] += 1
    for key, matches in primary_keys.items():
        if len(matches) > 1:
            counts[key[0]]["primary_rows_with_ambiguous_key"] += len(matches)

    totals = Counter()
    for language_counts in counts.values():
        totals.update(language_counts)
    return {"totals": totals, "languages": counts}


if __name__ == "__main__":
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("primary", type=Path)
    parser.add_argument("gold", type=Path)
    args = parser.parse_args()
    print(json.dumps(compare(args.primary, args.gold), indent=2, sort_keys=True))
