Datasets:
File size: 3,771 Bytes
71bdcd5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 | """Verify the public MAGAR-Bench files without reading Test gold labels."""
from __future__ import annotations
import argparse
import hashlib
import json
from collections import Counter
from pathlib import Path
EXPECTED_TASKS = {
"train": {"assembly_decision": 736, "atomic_action": 1764, "minimal_rework": 500},
"test": {"assembly_decision": 238, "atomic_action": 662, "minimal_rework": 100},
}
def read_jsonl(path: Path) -> list[dict]:
return [json.loads(line) for line in path.read_text(encoding="utf-8").splitlines() if line.strip()]
def file_hash(path: Path) -> str:
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 main() -> int:
parser = argparse.ArgumentParser()
parser.add_argument("--release-root", type=Path, default=Path(__file__).resolve().parents[1])
root = parser.parse_args().release_root.resolve()
train = read_jsonl(root / "data" / "train.jsonl")
test = read_jsonl(root / "data" / "test_inputs.jsonl")
errors: list[str] = []
if len(train) != 3000 or len(test) != 1000:
errors.append(f"row counts are train={len(train)}, test={len(test)}")
for split, rows in (("train", train), ("test", test)):
counts = Counter(row["metadata"]["task_family"] for row in rows)
if counts != Counter(EXPECTED_TASKS[split]):
errors.append(f"{split} task counts are {dict(counts)}")
train_ids = {row["id"] for row in train}
test_ids = {row["id"] for row in test}
if len(train_ids) != len(train) or len(test_ids) != len(test):
errors.append("duplicate sample ids")
if train_ids & test_ids:
errors.append("train/test id overlap")
train_families = {row["metadata"]["assembly_id"] for row in train}
test_families = {row["metadata"]["assembly_id"] for row in test}
if len(train_families) != 20 or len(test_families) != 6:
errors.append(f"family counts are train={len(train_families)}, test={len(test_families)}")
if train_families & test_families:
errors.append("train/test family overlap")
train_images = {image for row in train for image in row.get("images", [])}
test_images = {image for row in test for image in row.get("images", [])}
if train_images & test_images:
errors.append("train/test image overlap")
missing = [image for image in train_images | test_images if not (root / image).exists()]
if missing:
errors.append(f"missing image references: {len(missing)}")
if any(any(message.get("role") == "assistant" for message in row["messages"]) for row in test):
errors.append("test inputs contain assistant answers")
if any(any(key in row for key in ("target", "gold", "answer")) for row in test):
errors.append("test inputs contain target/gold/answer fields")
manifest = root / "MANIFEST.sha256"
if not manifest.exists():
errors.append("MANIFEST.sha256 is missing")
else:
for line in manifest.read_text(encoding="utf-8").splitlines():
if not line.strip():
continue
expected, relative = line.split(" ", 1)
target = root / relative
if not target.exists() or file_hash(target) != expected:
errors.append(f"manifest mismatch: {relative}")
result = {
"train_rows": len(train),
"test_rows": len(test),
"train_unique_images": len(train_images),
"test_unique_images": len(test_images),
"errors": errors,
}
print(json.dumps(result, ensure_ascii=False, indent=2))
return 1 if errors else 0
if __name__ == "__main__":
raise SystemExit(main())
|