scripts/validate_all.py

import argparse
import collections
import hashlib
import json
from pathlib import Path

import pyarrow.parquet as pq
from datasets import ClassLabel, Features, Value


LABELS = ["anger", "fear", "joy", "love", "sadness", "surprise", "neutral"]
LANGUAGES = {"de", "en", "es", "fr", "it"}
FEATURES = Features(
    {
        "id": Value("string"),
        "source_id": Value("string"),
        "text": Value("string"),
        "label": ClassLabel(names=LABELS),
        "label_name": Value("string"),
        "dataset": Value("string"),
        "license": Value("string"),
        "language": Value("string"),
        "is_label_conflict": Value("bool"),
        "label_variants": Value("int8"),
    }
)


def file_sha256(path):
    digest = hashlib.sha256()
    with path.open("rb") as stream:
        for chunk in iter(lambda: stream.read(1024 * 1024), b""):
            digest.update(chunk)
    return digest.hexdigest()


def parse_args():
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--reports",
        type=Path,
        nargs="+",
        default=[Path("reports/all_validation.json"), Path("reports/sample_validation.json")],
    )
    return parser.parse_args()


def validate(report_path):
    report = json.loads(report_path.read_text(encoding="utf-8"))
    ids = set()
    source_ids = set()
    rows = 0
    label_counts = collections.Counter()
    language_counts = collections.Counter()
    stratum_counts = collections.Counter()
    conflict_rows = 0

    for shard in report["shards"]:
        path = Path(shard["path"])
        if file_sha256(path) != shard["sha256"]:
            raise ValueError(f"checksum mismatch: {path}")
        parquet = pq.ParquetFile(path)
        if parquet.schema_arrow != FEATURES.arrow_schema:
            raise ValueError(f"schema mismatch: {path}")
        if parquet.metadata.num_rows != shard["rows"]:
            raise ValueError(f"row count mismatch: {path}")
        for batch in parquet.iter_batches(batch_size=65_536):
            columns = batch.to_pydict()
            for index, row_id in enumerate(columns["id"]):
                language = columns["language"][index]
                source_id = columns["source_id"][index]
                label = columns["label"][index]
                is_conflict = columns["is_label_conflict"][index]
                variants = columns["label_variants"][index]
                if row_id in ids:
                    raise ValueError(f"duplicate id: {row_id}")
                ids.add(row_id)
                source_ids.add(source_id)
                if language not in LANGUAGES:
                    raise ValueError(f"invalid language: {language}")
                if row_id != f"{source_id}_{language}":
                    raise ValueError(f"id/source mismatch: {row_id}")
                if not columns["text"][index].strip():
                    raise ValueError(f"empty text: {row_id}")
                if not 0 <= label < len(LABELS):
                    raise ValueError(f"invalid label: {row_id}")
                if columns["label_name"][index] != LABELS[label]:
                    raise ValueError(f"label name mismatch: {row_id}")
                if is_conflict != (variants > 1) or not 1 <= variants <= len(LABELS):
                    raise ValueError(f"conflict metadata mismatch: {row_id}")
                rows += 1
                conflict_rows += is_conflict
                label_counts[LABELS[label]] += 1
                language_counts[language] += 1
                stratum_counts[f"{language}/{LABELS[label]}"] += 1

    expected = report["counts"]
    checks = {
        "output_rows": rows,
        "output_unique_ids": len(ids),
        "retained_source_ids": len(source_ids),
        "label_conflict_rows": conflict_rows,
        "labels": dict(sorted(label_counts.items())),
        "languages": dict(sorted(language_counts.items())),
    }
    for key, actual in checks.items():
        if actual != expected[key]:
            raise ValueError(f"{report['config']} {key} mismatch: {actual} != {expected[key]}")
    if "language_label_strata" in expected:
        actual = dict(sorted(stratum_counts.items()))
        if actual != expected["language_label_strata"]:
            raise ValueError(f"{report['config']} language/label balance mismatch")
    return {
        "config": report["config"],
        "status": "valid",
        "rows": rows,
        "unique_ids": len(ids),
        "source_ids": len(source_ids),
        "shards": len(report["shards"]),
    }


def main():
    args = parse_args()
    print(json.dumps([validate(path) for path in args.reports], indent=2))


if __name__ == "__main__":
    main()