scripts/validate_dataset.py

#!/usr/bin/env python3
"""Validate JL-AgentBehavior-10K without third-party dependencies."""

from __future__ import annotations

import argparse
import hashlib
import json
from collections import Counter
from pathlib import Path
from typing import Any, Iterable


EXPECTED_SPLITS = {"train": 9000, "validation": 500, "test": 500}
EXPECTED_RECORD_TYPES = {"trajectory": 7000, "preference": 2000, "repair": 1000}
EXPECTED_BEHAVIORS = {
    "repository_grounding": 1500,
    "planning_decomposition": 1200,
    "tool_selection_execution": 1500,
    "bounded_code_editing": 2000,
    "test_verification": 1200,
    "failure_diagnosis_repair": 1200,
    "code_review_security": 800,
    "permission_scope_safety": 400,
    "final_reporting": 200,
}
EXPECTED_LANGUAGES = {
    "python": 2500,
    "typescript": 2500,
    "php": 1500,
    "java": 1500,
    "go": 1000,
    "rust": 400,
    "csharp": 300,
    "cpp": 300,
}
REQUIRED = {
    "id", "version", "split", "record_type", "primary_behavior", "secondary_behaviors",
    "language", "ecosystem", "task_type", "difficulty", "task", "environment",
    "behavioral_labels", "provenance", "quality", "supervision", "fingerprint",
}


def canonical_json(value: Any) -> str:
    return json.dumps(value, ensure_ascii=False, sort_keys=True, separators=(",", ":"))


def expected_fingerprint(record: dict[str, Any]) -> str:
    clean = dict(record)
    clean.pop("fingerprint", None)
    return hashlib.sha256(canonical_json(clean).encode("utf-8")).hexdigest()


def read_records(root: Path) -> Iterable[tuple[str, int, dict[str, Any]]]:
    for split in EXPECTED_SPLITS:
        path = root / "data" / f"{split}.jsonl"
        with path.open("r", encoding="utf-8") as handle:
            for line_number, line in enumerate(handle, 1):
                if line.strip():
                    yield split, line_number, json.loads(line)


def validate(root: Path) -> dict[str, Any]:
    errors: list[str] = []
    records: list[dict[str, Any]] = []
    ids: set[str] = set()
    fingerprints: set[str] = set()
    instructions: set[str] = set()
    split_repositories: dict[str, set[str]] = {split: set() for split in EXPECTED_SPLITS}

    for file_split, line_number, record in read_records(root):
        location = f"data/{file_split}.jsonl:{line_number}"
        records.append(record)
        missing = REQUIRED - record.keys()
        if missing:
            errors.append(f"{location}: missing fields {sorted(missing)}")
            continue
        if record["split"] != file_split:
            errors.append(f"{location}: split field does not match file")
        if record["id"] in ids:
            errors.append(f"{location}: duplicate id {record['id']}")
        ids.add(record["id"])
        if record["fingerprint"] in fingerprints:
            errors.append(f"{location}: duplicate fingerprint {record['fingerprint']}")
        fingerprints.add(record["fingerprint"])
        instruction = record["task"].get("instruction")
        if instruction in instructions:
            errors.append(f"{location}: duplicate task instruction")
        instructions.add(instruction)
        if record["fingerprint"] != expected_fingerprint(record):
            errors.append(f"{location}: fingerprint mismatch")
        if record["provenance"].get("execution_verified") is not False:
            errors.append(f"{location}: v1 synthetic record must not claim execution verification")
        if record["quality"].get("quality_tier") != "silver_structural":
            errors.append(f"{location}: unexpected quality tier")
        repo = record["environment"].get("repository_id")
        if not repo:
            errors.append(f"{location}: missing repository_id")
        else:
            split_repositories[file_split].add(repo)

    counters = {
        "split": Counter(record["split"] for record in records),
        "record_type": Counter(record["record_type"] for record in records),
        "primary_behavior": Counter(record["primary_behavior"] for record in records),
        "language": Counter(record["language"] for record in records),
    }
    expected = {
        "split": Counter(EXPECTED_SPLITS),
        "record_type": Counter(EXPECTED_RECORD_TYPES),
        "primary_behavior": Counter(EXPECTED_BEHAVIORS),
        "language": Counter(EXPECTED_LANGUAGES),
    }
    for name in counters:
        if counters[name] != expected[name]:
            errors.append(f"{name} distribution mismatch: {dict(counters[name])}")

    split_names = list(EXPECTED_SPLITS)
    for i, left in enumerate(split_names):
        for right in split_names[i + 1:]:
            overlap = split_repositories[left] & split_repositories[right]
            if overlap:
                errors.append(f"repository leakage between {left} and {right}: {sorted(overlap)[:5]}")

    return {
        "dataset": "JL-AgentBehavior-10K",
        "version": "1.0.0",
        "status": "passed" if not errors else "failed",
        "records_checked": len(records),
        "unique_ids": len(ids),
        "unique_fingerprints": len(fingerprints),
        "unique_instructions": len(instructions),
        "repository_family_overlap": False if not any("repository leakage" in e for e in errors) else True,
        "execution_verified_records": sum(r["provenance"]["execution_verified"] for r in records),
        "errors": errors,
        "counts": {name: dict(sorted(counter.items())) for name, counter in counters.items()},
    }


def main() -> None:
    parser = argparse.ArgumentParser()
    parser.add_argument("--root", type=Path, default=Path(__file__).resolve().parents[1])
    parser.add_argument("--report", type=Path)
    args = parser.parse_args()
    report = validate(args.root.resolve())
    rendered = json.dumps(report, indent=2, ensure_ascii=False) + "\n"
    if args.report:
        args.report.write_text(rendered, encoding="utf-8")
    print(rendered, end="")
    raise SystemExit(0 if report["status"] == "passed" else 1)


if __name__ == "__main__":
    main()