scripts/validate_dataset.py

from __future__ import annotations

import argparse
import json
import re
import sys
import tempfile
import zipfile
from collections import Counter, defaultdict
from pathlib import Path, PurePosixPath
from typing import Any

import duckdb
import pyarrow.parquet as pq
import yaml

from common import find_secret_patterns, sha256_file


EXPECTED_CONFIGS = ("papers", "archive_only", "chunks", "archives")
EXPECTED_SCHEMAS = {
    "papers": {
        "paper_id",
        "doi",
        "title",
        "authors",
        "date_published",
        "abstract",
        "keywords",
        "language",
        "genre",
        "canonical_url",
        "works_url",
        "tex_source",
        "archive_path",
        "archive_sha256",
        "tex_entry",
        "tex_sha256",
        "source_archive_paths",
        "source_archive_sha256s",
        "source_tex_entries",
        "source_tex_sha256s",
        "mapping_status",
        "content_status",
        "quality_flags",
    },
    "archive_only": {
        "record_id",
        "title",
        "authors",
        "date_raw",
        "language",
        "tex_source",
        "archive_path",
        "archive_sha256",
        "tex_entry",
        "tex_sha256",
        "mapping_status",
        "candidate_dois",
        "content_status",
        "duplicate_of_archive",
        "quality_flags",
    },
    "chunks": {
        "chunk_id",
        "paper_id",
        "doi",
        "title",
        "partition",
        "source_id",
        "source_archive_path",
        "tex_entry",
        "section_path",
        "section_title",
        "chunk_index",
        "char_start",
        "char_end",
        "chunk_tex",
        "chunk_text",
        "char_count",
        "quality_flags",
    },
    "archives": {
        "archive_id",
        "archive_path",
        "archive_filename",
        "archive_size",
        "archive_sha256",
        "mapped_dois",
        "candidate_dois",
        "mapping_status",
        "mapping_method",
        "mapping_score",
        "content_status",
        "primary_tex_entry",
        "primary_tex_sha256",
        "duplicate_of_archive",
        "entries",
        "quality_flags",
    },
}


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description="Validate a built paper TeX corpus.")
    parser.add_argument("--dataset", type=Path, required=True)
    parser.add_argument(
        "--skip-datasets",
        action="store_true",
        help="Skip the optional local load_dataset checks.",
    )
    parser.add_argument(
        "--compare",
        type=Path,
        help="Compare generated artifact hashes with another build directory.",
    )
    return parser.parse_args()


class Validation:
    def __init__(self) -> None:
        self.errors: list[str] = []
        self.warnings: list[str] = []
        self.checks: Counter[str] = Counter()

    def require(self, condition: bool, message: str) -> None:
        self.checks["assertions"] += 1
        if not condition:
            self.errors.append(message)

    def warn(self, condition: bool, message: str) -> None:
        if not condition:
            self.warnings.append(message)


def parquet_path(root: Path, config: str) -> Path:
    return root / "data" / config / "train-00000-of-00001.parquet"


def load_rows(root: Path, config: str) -> list[dict[str, Any]]:
    return pq.read_table(parquet_path(root, config)).to_pylist()


def parse_card_metadata(readme: str) -> dict[str, Any]:
    if not readme.startswith("---\n"):
        raise ValueError("README does not start with YAML front matter.")
    _, yaml_text, _ = readme.split("---", 2)
    return yaml.safe_load(yaml_text)


def verify_checksums(root: Path, validation: Validation) -> None:
    checksum_file = root / "checksums.sha256"
    validation.require(checksum_file.is_file(), "checksums.sha256 is missing.")
    if not checksum_file.is_file():
        return
    listed: set[str] = set()
    for line in checksum_file.read_text(encoding="utf-8").splitlines():
        if not line.strip():
            continue
        match = re.match(r"^([0-9a-f]{64})  (.+)$", line)
        validation.require(bool(match), f"Malformed checksum line: {line}")
        if not match:
            continue
        expected, relative = match.groups()
        listed.add(relative)
        path = root / Path(relative)
        validation.require(path.is_file(), f"Checksummed file is missing: {relative}")
        if path.is_file():
            validation.require(
                sha256_file(path) == expected,
                f"Checksum mismatch: {relative}",
            )
    expected_files: set[str] = set()
    excluded_parts = {".git", ".venv", "__pycache__", ".pytest_cache"}
    excluded_names = {
        "checksums.sha256",
        "publish-report.json",
        ".paper-tex-corpus-build",
    }
    for path in root.rglob("*"):
        if not path.is_file():
            continue
        relative = path.relative_to(root)
        if any(part in excluded_parts for part in relative.parts):
            continue
        if relative.name in excluded_names:
            continue
        expected_files.add(relative.as_posix())
    validation.require(
        listed == expected_files,
        "checksums.sha256 file set does not match the publishable artifact set.",
    )


def verify_card(root: Path, validation: Validation) -> dict[str, Any]:
    readme_path = root / "README.md"
    validation.require(readme_path.is_file(), "README.md is missing.")
    if not readme_path.is_file():
        return {}
    readme = readme_path.read_text(encoding="utf-8")
    try:
        metadata = parse_card_metadata(readme)
    except Exception as error:
        validation.errors.append(f"README YAML parsing failed: {error}")
        return {}
    validation.require(metadata.get("license") == "cc-by-4.0", "Card license mismatch.")
    configs = metadata.get("configs") or []
    config_names = [config.get("config_name") for config in configs]
    validation.require(
        config_names == list(EXPECTED_CONFIGS),
        f"Unexpected README config order/names: {config_names}",
    )
    defaults = [config for config in configs if config.get("default")]
    validation.require(
        len(defaults) == 1 and defaults[0].get("config_name") == "papers",
        "papers must be the only default config.",
    )
    validation.require(
        "not an evaluation benchmark" in readme,
        "Dataset limitations must state that this is not an evaluation benchmark.",
    )
    validation.require(
        "CC BY 4.0" in readme and "CITATION.cff" in readme,
        "README license/citation documentation is incomplete.",
    )
    return metadata


def verify_parquet(root: Path, report: dict[str, Any], validation: Validation) -> dict[str, list[dict[str, Any]]]:
    rows: dict[str, list[dict[str, Any]]] = {}
    for config in EXPECTED_CONFIGS:
        path = parquet_path(root, config)
        validation.require(path.is_file(), f"Missing Parquet for config {config}.")
        if not path.is_file():
            rows[config] = []
            continue
        parquet = pq.ParquetFile(path)
        actual_columns = set(parquet.schema_arrow.names)
        validation.require(
            actual_columns == EXPECTED_SCHEMAS[config],
            f"Schema mismatch for {config}: {sorted(actual_columns)}",
        )
        expected_count = report["counts"][config]
        validation.require(
            parquet.metadata.num_rows == expected_count,
            f"{config} row count differs from build-report.json.",
        )
        validation.require(
            parquet.metadata.num_row_groups > 0,
            f"{config} has no Parquet row groups.",
        )
        rows[config] = parquet.read().to_pylist()
    return rows


def verify_relations(
    root: Path,
    rows: dict[str, list[dict[str, Any]]],
    validation: Validation,
) -> None:
    papers = rows["papers"]
    archive_only = rows["archive_only"]
    chunks = rows["chunks"]
    archives = rows["archives"]

    paper_dois = {row["doi"] for row in papers}
    validation.require(len(papers) == 227, "papers must contain 227 scholarly records.")
    validation.require(
        len(paper_dois) == 227,
        "papers DOI values must be unique and non-missing.",
    )
    validation.require(
        all(doi.startswith("10.") for doi in paper_dois),
        "Every papers row must contain a DOI.",
    )
    validation.require(
        len(archives) == 250,
        "archives must contain exactly 250 rows.",
    )
    archive_paths = {row["archive_path"] for row in archives}
    validation.require(
        len(archive_paths) == 250,
        "Archive paths must be unique.",
    )
    archive_by_path = {row["archive_path"]: row for row in archives}

    for row in archives:
        path = root / Path(row["archive_path"])
        validation.require(path.is_file(), f"Raw archive is missing: {row['archive_path']}")
        if path.is_file():
            validation.require(
                path.stat().st_size == row["archive_size"],
                f"Raw archive size mismatch: {row['archive_path']}",
            )
            validation.require(
                sha256_file(path) == row["archive_sha256"],
                f"Raw archive hash mismatch: {row['archive_path']}",
            )
        validation.require(
            set(row["mapped_dois"]).issubset(paper_dois),
            f"Archive maps outside the papers DOI set: {row['archive_path']}",
        )

    expected_archive_only = {
        row["archive_path"] for row in archives if not row["mapped_dois"]
    }
    actual_archive_only = {row["archive_path"] for row in archive_only}
    validation.require(
        expected_archive_only == actual_archive_only,
        "archive_only rows do not exactly cover unmapped archives.",
    )

    for row in papers:
        validation.require(
            row["paper_id"] == row["doi"],
            f"paper_id must equal DOI: {row['paper_id']}",
        )
        validation.require(
            len(row["source_archive_paths"])
            == len(row["source_archive_sha256s"])
            == len(row["source_tex_entries"])
            == len(row["source_tex_sha256s"]),
            f"Source arrays differ in length for {row['doi']}.",
        )
        for archive_path in row["source_archive_paths"]:
            validation.require(
                archive_path in archive_by_path,
                f"Paper references unknown archive: {archive_path}",
            )
        if row["content_status"] == "valid":
            validation.require(
                bool(row["tex_source"] and row["tex_sha256"]),
                f"Valid paper has no TeX source: {row['doi']}",
            )
        if row["content_status"] == "metadata_only":
            validation.require(
                not row["tex_source"] and not row["archive_path"],
                f"metadata_only paper unexpectedly has a source: {row['doi']}",
            )

    chunk_ids = [row["chunk_id"] for row in chunks]
    validation.require(
        len(chunk_ids) == len(set(chunk_ids)),
        "chunk_id values must be unique.",
    )
    chunk_sources = defaultdict(list)
    for row in chunks:
        chunk_sources[row["source_id"]].append(row)
        validation.require(
            row["source_archive_path"] in archive_by_path,
            f"Chunk references unknown archive: {row['chunk_id']}",
        )
        validation.require(
            row["char_end"] > row["char_start"],
            f"Chunk has invalid character span: {row['chunk_id']}",
        )
        validation.require(
            row["char_count"] == len(row["chunk_tex"]),
            f"Chunk char_count mismatch: {row['chunk_id']}",
        )
        validation.require(
            row["partition"] in {"papers", "archive_only"},
            f"Chunk partition is invalid: {row['chunk_id']}",
        )
        if row["doi"]:
            validation.require(
                row["doi"] in paper_dois,
                f"Chunk DOI is absent from papers: {row['chunk_id']}",
            )
    for source_id, source_rows in chunk_sources.items():
        indexes = [row["chunk_index"] for row in source_rows]
        validation.require(
            indexes == list(range(len(indexes))),
            f"Chunk indexes are not contiguous for {source_id}.",
        )


def verify_zip_security(root: Path, archive_rows: list[dict[str, Any]], validation: Validation) -> None:
    for row in archive_rows:
        path = root / Path(row["archive_path"])
        if not path.is_file():
            continue
        try:
            with zipfile.ZipFile(path) as archive:
                validation.require(
                    archive.testzip() is None,
                    f"ZIP CRC validation failed: {row['archive_path']}",
                )
                for info in archive.infolist():
                    member = PurePosixPath(info.filename.replace("\\", "/"))
                    validation.require(
                        not member.is_absolute()
                        and ".." not in member.parts
                        and not re.match(r"^[A-Za-z]:", info.filename),
                        f"Unsafe ZIP path: {row['archive_path']}::{info.filename}",
                    )
                    if info.is_dir() or not info.filename.lower().endswith(".tex"):
                        continue
                    text = archive.read(info).decode("utf-8", errors="replace")
                    validation.require(
                        "\ufffd" not in text,
                        f"UTF-8 replacement character in {row['archive_path']}::{info.filename}",
                    )
                    validation.require(
                        not find_secret_patterns(text),
                        f"Secret-like pattern in {row['archive_path']}::{info.filename}",
                    )
        except zipfile.BadZipFile as error:
            validation.errors.append(f"Broken ZIP {row['archive_path']}: {error}")


def verify_duckdb(root: Path, report: dict[str, Any], validation: Validation) -> None:
    connection = duckdb.connect(":memory:")
    try:
        for config in EXPECTED_CONFIGS:
            path = parquet_path(root, config).as_posix().replace("'", "''")
            result = connection.execute(
                f"SELECT count(*) FROM read_parquet('{path}')"
            ).fetchone()[0]
            validation.require(
                result == report["counts"][config],
                f"DuckDB count mismatch for {config}.",
            )
    finally:
        connection.close()


def verify_datasets(root: Path, report: dict[str, Any], validation: Validation) -> None:
    try:
        from datasets import load_dataset
    except ImportError:
        validation.errors.append("datasets package is unavailable.")
        return
    with tempfile.TemporaryDirectory(prefix="paper-tex-corpus-datasets-") as cache:
        for config in EXPECTED_CONFIGS:
            try:
                dataset = load_dataset(
                    str(root),
                    config,
                    split="train",
                    cache_dir=str(Path(cache) / config),
                    download_mode="force_redownload",
                )
            except Exception as error:
                validation.errors.append(
                    f"datasets.load_dataset failed for {config}: {error}"
                )
                continue
            validation.require(
                len(dataset) == report["counts"][config],
                f"datasets row count mismatch for {config}.",
            )


def artifact_hashes(root: Path) -> dict[str, str]:
    selected: dict[str, str] = {}
    for prefix in ("data", "raw", "metadata"):
        base = root / prefix
        for path in sorted(item for item in base.rglob("*") if item.is_file()):
            relative = path.relative_to(root).as_posix()
            selected[relative] = sha256_file(path)
    for name in ("README.md", "LICENSE", "CITATION.cff", "build-report.json"):
        selected[name] = sha256_file(root / name)
    return selected


def verify_comparison(root: Path, other: Path, validation: Validation) -> None:
    other = other.resolve()
    validation.require(other.is_dir(), f"Comparison build is missing: {other}")
    if not other.is_dir():
        return
    validation.require(
        artifact_hashes(root) == artifact_hashes(other),
        "Deterministic rebuild comparison failed.",
    )


def main() -> None:
    args = parse_args()
    root = args.dataset.resolve()
    validation = Validation()
    validation.require(root.is_dir(), f"Dataset directory is missing: {root}")
    validation.require(
        (root / ".paper-tex-corpus-build").is_file(),
        "Dataset build sentinel is missing.",
    )
    report_path = root / "build-report.json"
    validation.require(report_path.is_file(), "build-report.json is missing.")
    if not report_path.is_file():
        print(json.dumps({"status": "failed", "errors": validation.errors}, indent=2))
        raise SystemExit(1)
    report = json.loads(report_path.read_text(encoding="utf-8"))
    verify_checksums(root, validation)
    verify_card(root, validation)
    rows = verify_parquet(root, report, validation)
    verify_relations(root, rows, validation)
    verify_zip_security(root, rows["archives"], validation)
    verify_duckdb(root, report, validation)
    if not args.skip_datasets:
        verify_datasets(root, report, validation)
    if args.compare:
        verify_comparison(root, args.compare, validation)
    result = {
        "status": "passed" if not validation.errors else "failed",
        "assertions": validation.checks["assertions"],
        "errors": validation.errors,
        "warnings": validation.warnings,
        "counts": report["counts"],
    }
    print(json.dumps(result, ensure_ascii=False, indent=2, sort_keys=True))
    if validation.errors:
        raise SystemExit(1)


if __name__ == "__main__":
    main()