Dataset / curation /v3 /scripts /reverify_export.py
dsfasdfaserwer's picture
Add verified Python TACO v3 pilot part 1/5 (2 rows)
22c9027 verified
Raw History Blame Contribute Delete
3.04 kB
"""Replay a downloaded v3 export. Requires Linux chroot/seccomp setup privileges.
Example (from the downloaded HF repository root):
python curation/v3/scripts/reverify_export.py --repo-root .
Only the messages field is counted as training input. No source-provided scripts
are imported or executed in the host process. Candidate Python runs in the
restricted worker, never with an HF token. No credentials are read by this tool.
"""
import argparse
import gzip
import hashlib
import json
from pathlib import Path
import jsonschema
from transformers import AutoTokenizer
from verify import self_test, static_check, verify_suite, sha
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--repo-root', type=Path, required=True)
parser.add_argument('--tokenizer-dir', type=Path, help='Optional local copy of the pinned tokenizer')
args = parser.parse_args()
root = args.repo_root.resolve()
self_test()
meta = json.loads((root/'curation/v3/tokenization.json').read_text())
if args.tokenizer_dir:
tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_dir, local_files_only=True, trust_remote_code=False)
else:
tokenizer = AutoTokenizer.from_pretrained(meta['model'],revision=meta['revision'],trust_remote_code=False,token=False)
if sha(tokenizer.chat_template) != meta['chat_template_sha256']:
raise RuntimeError('Wrong tokenizer/chat template')
validator = jsonschema.Draft202012Validator(json.loads((root/'schemas/python_verified_v3.schema.json').read_text()))
rows = total = 0
paths = sorted(root.glob('data/16k_context/taco_v3_pilot_*.jsonl')) + sorted(root.glob('data/above_16k/taco_v3_pilot_*.jsonl'))
for path in paths:
for line in path.read_text().splitlines():
if not line.strip(): continue
row = json.loads(line); validator.validate(row); static_check(row['code'])
count = len(tokenizer.apply_chat_template(row['messages'], tokenize=True, add_generation_prompt=False))
if count != row['token_counts']['training_total']:
raise RuntimeError('Stale token count: '+row['id'])
if ('16k_context' if count<=16384 else 'above_16k') != row['context_bucket']:
raise RuntimeError('Wrong bucket')
artifact = root/row['tests']['artifact_path']
if hashlib.sha256(artifact.read_bytes()).hexdigest() != row['tests']['artifact_sha256']:
raise RuntimeError('Test artifact hash mismatch')
suite = json.loads(gzip.decompress(artifact.read_bytes()))
result = verify_suite(row['code'],suite['entry_point'],suite['cases'])
if result['failed']:
raise RuntimeError(json.dumps(result,indent=2))
rows+=1; total+=result['total']
print(row['id'],f"{result['passed']}/{result['total']}",count,'tokens',flush=True)
print(f'Reverified {rows} Python rows; {total} successful test-case executions.')
if __name__ == '__main__':
main()