"""Use the unchanged formal B1 visual encoder; publish only complete QA assets.""" import argparse import json from pathlib import Path import sys from types import SimpleNamespace from common import ROOT,DEFAULT_RUN,sha256,write_json from external_visual import DEFAULT sys.path[:0]=[str(ROOT),str(ROOT/'src'),str(ROOT/'scripts/data')] def features(root,rank,world): from prepare_hfd_video_assets import encode plan=json.loads((root/'speech/plan.json').read_text()) media={r['video_sha256']:r for r in plan['rows']} rows=[dict(asset_id=sha,video_path=r['video_path'],video_sha256=sha, duration_seconds=min(r['duration_seconds'],180.)) for sha,r in sorted(media.items())] # Only stop encoding unneeded suffix; never move the query, speed up video, # or shorten the observed prefix. Sessions must fit the encoded duration. output=root/'features' write_json(output/'admission/split.json',dict(rows=rows)) config=json.loads((DEFAULT_RUN/'run_config.json').read_text()) encode(SimpleNamespace(root=output,base_model=Path(config['model']['model_name_or_path']), selected_before_split=False,world=world,rank=rank,limit=0,packet_batch=1)) def publish(root,output,allowlist=None): from expand_hfd_video_voices import VOICES,translation_path,audio_path,canonical plan=json.loads((root/'speech/plan.json').read_text()) ids={r['question']:r['id'] for r in plan['questions']} manual_path=root/'manual_review.json' manual=json.loads(manual_path.read_text())['rows'] if manual_path.exists() else [] veto={r['id'] for r in manual if not r['accepted']} inspected=None if allowlist: inspection=json.loads(allowlist.read_text()) if inspection.get('complete') is not True: raise ValueError('Direct visual inspection unfinished') inspected={r['id']:r for r in inspection['rows']} if len(inspected)!=len(inspection['rows']):raise ValueError('Duplicate direct inspection IDs') rows=[];rejections=[] for row in plan['rows']: if inspected is not None and not inspected.get(row['id'],{}).get('accepted'): rejections.append(dict(id=row['id'],reason='direct_inspection_not_approved'));continue if row['id'] in veto: rejections.append(dict(id=row['id'],reason='manual_visual_veto'));continue qtranslation=translation_path(root/'speech','question',row['question']) atranslation=translation_path(root/'speech','answer:'+row['question'],row['answer']) if not qtranslation.exists() or not atranslation.exists(): raise ValueError('Translation review unfinished') if not all(json.loads(p.read_text())['accepted'] for p in (qtranslation,atranslation)): rejections.append(dict(id=row['id'],reason='translation_not_verified'));continue fpath=root/'features/admission/feature_receipts'/(row['video_sha256']+'.json') if not fpath.exists():raise ValueError('Feature extraction not finished: '+str(fpath)) feature=json.loads(fpath.read_text()) if feature.get('rejected'): rejections.append(dict(id=row['id'],reason='feature_decode_rejected'));continue if sha256(feature['path'])!=feature['sha256']:raise ValueError('Feature file changed') qid=ids[row['question']];variants=[] for voice in VOICES: wave=audio_path(root/'speech',qid,voice);receipt=wave.with_suffix('.json') if not receipt.exists():raise ValueError('TTS/ASR unfinished: '+str(receipt)) ar=json.loads(receipt.read_text()) if not ar.get('accepted'):break if (ar['wave_sha256']!=sha256(wave) or not ar['asr'].get('independent_asr') or canonical(ar['asr']['heard'],voice[-2:])!=canonical(ar['text'],voice[-2:])): raise ValueError('Independent ASR receipt mismatch') question,answer=row['question'],row['answer'] if voice.endswith('_zh'): q=json.loads(translation_path(root/'speech','question',question).read_text()) ans=json.loads(translation_path(root/'speech','answer:'+question,answer).read_text()) if not q['accepted'] or not ans['accepted']:break question,answer=q['translation'],ans['translation'] if ar['text']!=question:raise ValueError('Question text changed after TTS') variants.append(dict(asset_id=row['id'],external_id=row['id'],voice=voice,split='val', data_domain='OVO-Bench',question=question,answer=answer, original_question=row['question'],original_answer=row['answer'], question_time=row['question_time'],options=row['options'],correct_option=row['correct_option'], video_path=row['video_path'],video_sha256=row['video_sha256'], duration_seconds=min(row['duration_seconds'],180.), original_video_duration_seconds=row['duration_seconds'], visual_features_path=feature['path'],visual_features_sha256=feature['sha256'], visual_receipt=row['visual_receipt'],visual_receipt_sha256=sha256(row['visual_receipt']), question_audio_path=str(wave),question_audio_sha256=ar['wave_sha256'],asr_receipt=str(receipt))) if len(variants)!=8: rejections.append(dict(id=row['id'],reason='incomplete_verified_eight_voices'));continue rows.extend(variants) if allowlist: for variant in variants: variant.update(direct_inspection_path=str(allowlist.resolve()), direct_inspection_sha256=sha256(allowlist)) if not rows:raise ValueError('No complete reviewed external QA') payload=dict(format='hfd_bench_external_assets_v1',assets=rows,rejections=rejections, provenance=dict(source='OVO-Bench',speech_plan_sha256=sha256(root/'speech/plan.json'), split_audit_sha256=sha256(root/'split_audit.json'),original_options_preserved=True, manual_review_sha256=sha256(manual_path) if manual_path.exists() else None, direct_inspection_sha256=sha256(allowlist) if allowlist else None, protocol='HFD spoken open QA adaptation; question END at original timestamp; no history crop', training_data_modified=False)) if output.exists() and json.loads(output.read_text())!=payload:raise ValueError('Refusing to overwrite a changed published manifest') write_json(output,payload) print('PUBLISHED',len(rows)//8,'questions',len(rows),'voices',output,flush=True) if __name__=='__main__': p=argparse.ArgumentParser(description=__doc__) p.add_argument('command',choices=['features','publish']) p.add_argument('--root',type=Path,default=DEFAULT) p.add_argument('--output',type=Path,default=ROOT/'benchmark/outputs/external_visual/overlay_assets.json') p.add_argument('--manual-allowlist',type=Path,help='Completed direct inspection; only explicit accepted IDs admitted') p.add_argument('--rank',type=int,default=0);p.add_argument('--world',type=int,default=1) a=p.parse_args() if a.command=='features':features(a.root,a.rank,a.world) else:publish(a.root,a.output,a.manual_allowlist)