"""Benchmark-only acquisition and provenance. No training dataset is modified. `fetch --media` resumes official OVO source archives (not redundant chunked archives). `inventory` is a structural filter, NOT a visual quality approval. """ import argparse from collections import Counter from concurrent.futures import ThreadPoolExecutor import io import json from pathlib import Path, PurePosixPath import shutil import tarfile import time import requests from common import sha256, write_json DEFAULT = Path('/mnt/workspace/hfd_bench_external_261013') OVO = 'https://raw.githubusercontent.com/JoeLeelyf/OVO-Bench/main/data/ovo_bench_new.json' TASKS = {'EPM', 'HLD', 'ASI', 'STU', 'OJR', 'ATR', 'ACR', 'OCR'} DEFERRED_ROOTS = ('Ego4D/', 'OpenEQA/') # Do not infer signed access from a mirror. def fetch(root, media=False): from huggingface_hub import hf_hub_download root.mkdir(parents=True, exist_ok=True) path = root/'annotations/ovo.json' if not path.exists(): response = requests.get(OVO, timeout=60); response.raise_for_status() write_json(path, response.json()) write_json(path.with_suffix('.provenance.json'), dict(url=OVO, sha256=sha256(path), retrieved_at=time.time())) jobs = [('JoeLeelyf/OVO-Bench', 'README.md'), ('stdKonjac/LiveSports-3K', 'README.md'), ('stdKonjac/LiveSports-3K', 'LiveSports-3K-CC.json'), ('stdKonjac/LiveSports-3K', 'LiveSports-3K-QA.json'), ('Hothan/Omni-DuplexEval', 'README.md')] if media: jobs += [('JoeLeelyf/OVO-Bench', 'src_videos.tar.part'+s) for s in ('aa', 'ab', 'ac', 'ad', 'ae')] revisions = {} for repo, _ in jobs: if repo not in revisions: response = requests.get('https://huggingface.co/api/datasets/'+repo, timeout=60) response.raise_for_status(); revisions[repo] = response.json()['sha'] frozen = root/'revisions.json' if frozen.exists(): revisions = json.loads(frozen.read_text()) else: write_json(frozen, revisions) def one(job): repo, name = job out = hf_hub_download(repo, name, repo_type='dataset', revision=revisions[repo], local_dir=root/'downloads'/repo.split('/')[-1]) print('DOWNLOADED', repo, name, Path(out).stat().st_size, flush=True) return out with ThreadPoolExecutor(max_workers=4) as pool: list(pool.map(one, jobs)) def exclusion(row): if row['task'] not in TASKS: return 'anticipation_or_proactive_task_not_prefix_qa' if row['video'].startswith(DEFERRED_ROOTS): return 'source_access_not_confirmed' if row['video'].startswith('YouTube_Games/'): return 'game_visual_not_primary_robot_scenario' if not 5 <= float(row['realtime']) <= 120: return 'query_outside_current_session_length_budget' if not row.get('question', '').strip() or not row.get('answer', '').strip(): return 'empty_text' if not 0 <= int(row['gt']) < len(row['options']): return 'invalid_option_index' # Do not require answer == option: official full-sentence references often # paraphrase the correct option. Preserve both, never overwrite either. return None def inventory(root): rows = json.loads((root/'annotations/ovo.json').read_text()) accepted, rejected = [], [] for row in rows: reason = exclusion(row) if reason: rejected.append(dict(id=row['id'], reason=reason)); continue accepted.append(dict(id=f"ovo_{row['id']:04d}", source='OVO-Bench', original=row, video_relative=row['video'], question_time=float(row['realtime']), question=row['question'].strip(), answer=row['answer'].strip(), options=row['options'], correct_option=int(row['gt']), visual_review_status='pending', protocol='HFD spoken open QA adaptation; original options retained for later evaluation', # No time relocation, video speed change, loop or history crop. video_start_seconds=0.0)) write_json(root/'candidates.json', accepted) write_json(root/'structural_rejections.json', rejected) summary = dict(official_records=len(rows), structurally_eligible=len(accepted), unique_videos=len({r['video_relative'] for r in accepted}), tasks=dict(Counter(r['original']['task'] for r in accepted)), rejections=dict(Counter(r['reason'] for r in rejected)), visually_approved=0, training_ready=False, other_sources={ 'Omni-DuplexEval': 'Deferred: continuous descriptions / proactive events need a different response protocol, not fabricated one-shot QA.', 'LiveSports-3K-CC': 'Not used as factual QA: ASR commentary includes nonvisual biography/context.', 'LiveSports-3K-QA': 'Potential alternative, not CC. Audio-dependent questions require filtering; not silently substituted.'}) write_json(root/'inventory.json', summary) print(json.dumps(summary, ensure_ascii=False, indent=2), flush=True) class JoinedReader(io.RawIOBase): """Read split tar in order without writing a second 46 GB joined archive.""" def __init__(self, paths): self.paths = iter(paths); self.current = None def readable(self): return True def readinto(self, buffer): if self.current is None: path = next(self.paths, None) if path is None: return 0 self.current = path.open('rb') count = self.current.readinto(buffer) if count: return count self.current.close(); self.current = None return self.readinto(buffer) def close(self): if self.current is not None: self.current.close() super().close() def extract(root): wanted = {r['video_relative'] for r in json.loads((root/'candidates.json').read_text())} parts = [root/'downloads/OVO-Bench'/('src_videos.tar.part'+s) for s in ('aa','ab','ac','ad','ae')] found = set(); sizes = {} with JoinedReader(parts) as joined, io.BufferedReader(joined, 8*1024*1024) as stream: with tarfile.open(fileobj=stream, mode='r|') as archive: for member in archive: name = PurePosixPath(member.name) if name.is_absolute() or '..' in name.parts or not member.isfile(): continue matches = [str(PurePosixPath(*name.parts[i:])) for i in range(len(name.parts))] key = next((s for s in matches if s in wanted), None) if key is None: continue dest = root/'media/OVO-Bench'/key if dest.is_file() and dest.stat().st_size == member.size: found.add(key); sizes[key] = member.size; continue dest.parent.mkdir(parents=True, exist_ok=True) with archive.extractfile(member) as source, dest.with_suffix('.partial').open('wb') as out: shutil.copyfileobj(source, out, 8*1024*1024) dest.with_suffix('.partial').replace(dest) found.add(key); sizes[key] = member.size print('EXTRACTED',len(found),'/',len(wanted),key,flush=True) write_json(root/'extraction.json', dict(found=sorted(found), missing=sorted(wanted-found), bytes=sum(sizes.values()))) if wanted-found: raise RuntimeError(f'Missing {len(wanted-found)} requested videos; inspect extraction.json') if __name__ == '__main__': p = argparse.ArgumentParser(description=__doc__) p.add_argument('command', choices=['fetch','inventory','extract']) p.add_argument('--root', type=Path, default=DEFAULT) p.add_argument('--media', action='store_true') a = p.parse_args() if a.command == 'fetch': fetch(a.root, a.media) elif a.command == 'inventory': inventory(a.root) else: extract(a.root)