scripts/audit_stack_v3_shard.py

#!/usr/bin/env python3
"""Audit one Stack v3 TRAIN shard using only nested metadata columns."""

from __future__ import annotations

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

import pyarrow.parquet as pq


def sha256(path: Path) -> str:
    digest = hashlib.sha256()
    with path.open("rb") as handle:
        for chunk in iter(lambda: handle.read(8 * 1024 * 1024), b""):
            digest.update(chunk)
    return digest.hexdigest()


def main() -> None:
    parser = argparse.ArgumentParser()
    parser.add_argument("--input", type=Path, required=True)
    parser.add_argument("--panel", type=Path, required=True)
    parser.add_argument("--revision", required=True)
    parser.add_argument("--generated-at", required=True)
    parser.add_argument("--output", type=Path, required=True)
    args = parser.parse_args()

    panel = set(json.loads(args.panel.read_text())["matrix_languages"])
    counts: dict[str, Counter[str]] = defaultdict(Counter)
    eligible_repos: dict[str, set[str]] = defaultdict(set)
    license_counts: dict[str, Counter[str]] = defaultdict(Counter)
    license_combo_counts: dict[str, Counter[tuple[str, ...]]] = defaultdict(Counter)
    content_ids: Counter[str] = Counter()
    repo_rows = 0
    mixed_panel_language_repos = 0
    declared_files = 0
    observed_files = 0
    inconsistent_num_files_rows = 0

    parquet = pq.ParquetFile(args.input)
    columns = [
        "repo_path",
        "num_files",
        "files.list.element.content_id",
        "files.list.element.size_bytes",
        "files.list.element.language",
        "files.list.element.is_vendor",
        "files.list.element.license_type",
        "files.list.element.detected_licenses",
    ]
    for batch in parquet.iter_batches(batch_size=128, columns=columns):
        for repo in batch.to_pylist():
            repo_rows += 1
            files = repo["files"] or []
            declared_files += repo["num_files"]
            observed_files += len(files)
            if repo["num_files"] != len(files):
                inconsistent_num_files_rows += 1
            repo_panel_languages = set()
            for file in files:
                language = file["language"]
                size = file["size_bytes"] or 0
                if language not in panel:
                    continue
                repo_panel_languages.add(language)
                counts[language]["files"] += 1
                counts[language]["bytes"] += size
                license_type = file["license_type"] or "null"
                counts[language][f"license_type_{license_type}_files"] += 1
                counts[language][f"license_type_{license_type}_bytes"] += size
                if file["is_vendor"]:
                    counts[language]["vendor_files"] += 1
                    counts[language]["vendor_bytes"] += size
                licenses = tuple(sorted(file["detected_licenses"] or []))
                for license_id in licenses:
                    license_counts[language][license_id] += 1
                license_combo_counts[language][licenses] += 1
                if license_type == "permissive" and not file["is_vendor"]:
                    counts[language]["eligible_files"] += 1
                    counts[language]["eligible_bytes"] += size
                    eligible_repos[language].add(repo["repo_path"])
                    if file["content_id"]:
                        content_ids[file["content_id"]] += 1
            if len(repo_panel_languages) > 1:
                mixed_panel_language_repos += 1

    language_rows = {}
    for language in sorted(panel):
        row = dict(sorted(counts[language].items()))
        row["eligible_repositories"] = len(eligible_repos[language])
        row["detected_license_counts"] = dict(
            sorted(license_counts[language].items(), key=lambda item: (-item[1], item[0]))
        )
        row["detected_license_combination_counts"] = {
            ";".join(combo) if combo else "<none>": count
            for combo, count in sorted(
                license_combo_counts[language].items(),
                key=lambda item: (-item[1], item[0]),
            )
        }
        language_rows[language] = row

    result = {
        "schema_version": "1.0.0",
        "source_id": "hf_stack_v3_train",
        "artifact_revision": args.revision,
        "generated_at": args.generated_at,
        "input": {
            "filename": args.input.name,
            "size_bytes": args.input.stat().st_size,
            "sha256": sha256(args.input),
            "parquet_rows": parquet.metadata.num_rows,
            "parquet_row_groups": parquet.metadata.num_row_groups,
        },
        "audit_scope": "metadata-only pilot; file content was not read or republished",
        "warning": "A single shard is a pipeline validation artifact, not a statistically guaranteed representative sample and must not be extrapolated to final corpus totals.",
        "repository_rows": repo_rows,
        "declared_files": declared_files,
        "observed_files": observed_files,
        "rows_with_num_files_mismatch": inconsistent_num_files_rows,
        "repositories_with_multiple_panel_languages": mixed_panel_language_repos,
        "eligible_definition": "matrix language AND license_type=permissive AND is_vendor=false; detected-license allowlist and other project filters are still pending",
        "eligible_content_ids": len(content_ids),
        "repeated_eligible_content_ids": sum(count > 1 for count in content_ids.values()),
        "maximum_eligible_content_id_multiplicity": max(content_ids.values(), default=0),
        "languages": language_rows,
    }
    args.output.parent.mkdir(parents=True, exist_ok=True)
    args.output.write_text(json.dumps(result, indent=2, sort_keys=True) + "\n")


if __name__ == "__main__":
    main()