Datasets:
File size: 38,451 Bytes
ce83ac1 | 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 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 | #!/usr/bin/env python3
"""Export current, unwrapped Code2Video task inputs; no media or API calls."""
from __future__ import annotations
import argparse
import ast
import csv
import hashlib
import json
import math
import re
import sys
import zipfile
from collections import Counter
from concurrent.futures import ThreadPoolExecutor
from datetime import datetime, timezone
from pathlib import Path
TASKS = ('t2v', 'i2v', 'v2v')
VERSION = '1.2'
def dumps(value):
return json.dumps(value, ensure_ascii=False, sort_keys=True)
def read_json(path, required=False):
if not path.is_file():
if required:
raise FileNotFoundError(path)
return {}
return json.loads(path.read_text(encoding='utf-8'))
def sha_text(value):
return hashlib.sha256(value.encode('utf-8')).hexdigest()
def number(value):
if value in (None, '') or isinstance(value, bool):
return None
if isinstance(value, (int, float)):
return float(value) if math.isfinite(value) else None
text = re.sub(r'\s*(?:seconds?|secs?|s)\s*$', '', str(value).strip(), flags=re.I)
if not re.fullmatch(r'-?\d+(?:\.\d+)?(?::\d+(?:\.\d+)?){0,2}', text):
return None
parts = text.split(':')
result = 0.0
try:
for part in parts:
result = result * 60 + float(part)
return result if math.isfinite(result) else None
except ValueError:
return None
def urls(value):
if isinstance(value, list):
return list(dict.fromkeys(u for v in value for u in urls(v)))
if not isinstance(value, str):
return []
return list(dict.fromkeys(re.findall(r'https?://[^\s|<>"\[\]]+', value)))
def read_csv(path):
if not path.is_file():
return []
with path.open(encoding='utf-8-sig', newline='') as stream:
return list(csv.DictReader(stream))
def write_json(path, value):
path.parent.mkdir(parents=True, exist_ok=True)
path.write_text(json.dumps(value, ensure_ascii=False, indent=2) + '\n', encoding='utf-8')
def write_csv(path, rows):
if not rows:
path.write_text('', encoding='utf-8-sig')
return
fields = list(dict.fromkeys(k for row in rows for k in row))
with path.open('w', encoding='utf-8-sig', newline='') as stream:
writer = csv.DictWriter(stream, fieldnames=fields)
writer.writeheader()
writer.writerows({k: dumps(v) if isinstance(v, (list, dict)) else v
for k, v in row.items()} for row in rows)
def taxonomy(project):
"""Read literal definitions, without importing or changing project modules."""
paths = [project / 'code_video_pilot/scripts/source_quality/video_taxonomy.py',
project / 'code_video_pilot/scripts/video_prompt/taxonomy.py']
result = {}
for path in paths:
for node in ast.parse(path.read_text()).body:
if isinstance(node, ast.Assign) and isinstance(node.value, ast.Dict):
for target in node.targets:
if isinstance(target, ast.Name) and target.id in {'FAMILY_NAMES', 'SUBTYPE_NAMES', 'SUBTYPE_FOCUS'}:
result[target.id] = ast.literal_eval(node.value)
for field in ('FAMILY_NAMES', 'SUBTYPE_NAMES'):
result[field] = {key: value.rsplit(' / ', 1)[-1]
for key, value in result.get(field, {}).items()}
return result
def collect_sample(job):
collection, folder, source_number = job
meta = read_json(folder / 'metadata.json', required=True)
annotation = read_json(folder / 'annotation.json')
task = read_json(folder / 'task.json') if collection == 'generate' else {}
seed = read_json(folder / 'seed.json') if collection == 'generate' else {}
payloads = {setting: read_json(folder / f'{setting}.json') for setting in TASKS
if (folder / f'{setting}.json').is_file()}
return {'collection': collection, 'original_id': folder.name, 'source_number': source_number,
'folder': str(folder), 'metadata': meta, 'annotation': annotation,
'task': task, 'seed': seed, 'payloads': payloads}
def list_samples(input_root):
indexed = read_csv(input_root / 'index.csv')
order = {(r['collection'], r['sample_id']): i for i, r in enumerate(indexed)}
samples = []
for collection in ('reproduce', 'generate'):
directory = input_root / f'code2video-{collection if collection == "reproduce" else "gen"}'
folders = [p for p in directory.iterdir() if p.is_dir() and not p.name.startswith(('_', '.'))]
folders.sort(key=lambda p: (order.get((collection, p.name), len(order)), p.name))
samples.extend((collection, p, 0) for p in folders)
return [(collection, folder, i + 1) for i, (collection, folder, _) in enumerate(samples)]
def source_manifest(project):
path = project / 'code_video_pilot/data/quality/high_score_gallery_20261004/reviewed_final/prompt_input_manifest.csv'
return {r['sample_id']: r for r in read_csv(path)}
def first_nonempty(*values):
return next((v for v in values if v not in (None, '', [], {})), None)
def label_ids(meta, source):
tax = meta.get('taxonomy') or {}
candidates = first_nonempty(meta.get('final_category_ids'), meta.get('subtype_ids'),
[tax['subtype_id']] if tax.get('subtype_id') else None,
re.findall(r'\b\d+\.\d+\b', source.get('prompt_category_ids', ''))) or []
subtypes = list(dict.fromkeys(str(x) for x in candidates if '.' in str(x)))
return list(dict.fromkeys(s.split('.')[0] for s in subtypes)), subtypes
def as_strings(value):
if value is None:
return []
if not isinstance(value, list):
value = [value]
return [v if isinstance(v, str) else dumps(v) for v in value]
def public_metadata(value):
"""Drop machine-local paths; originals remain in the private snapshot."""
if isinstance(value, dict):
return {k: public_metadata(v) for k, v in value.items()
if not k.endswith('_path') and k not in {'path', 'source_reference_path'}}
if isinstance(value, list):
return [public_metadata(v) for v in value]
if isinstance(value, str) and value.startswith(('/mnt/', '/Users/', 'file://')):
return None
return value
def make_records(sample, manifest, tax, reference_sources=None):
meta = sample['metadata']; task = sample['task']; seed = sample['seed']
source = manifest.get(sample['original_id'], {}) if sample['collection'] == 'reproduce' else {}
annotation = first_nonempty(sample['annotation'].get('source_annotation'),
task.get('source_annotation'), meta.get('source_annotation'), {}) or {}
fields = meta.get('fields') or {}
summary = first_nonempty(annotation.get('summary'), fields.get('brief_summary'))
categories, subtypes = label_ids(meta, source)
difficulty = meta.get('difficulty') or {}
C = first_nonempty(fields.get('construction_complexity'), difficulty.get('target_C'))
D = first_nonempty(fields.get('dynamic_complexity'), difficulty.get('target_D'))
S = first_nonempty(fields.get('dependency_span'), difficulty.get('target_S'))
group_id = f'C2VC-{sample["source_number"]:04d}'
subjects = annotation.get('primary_subject', [])
events = annotation.get('events', [])
motion = annotation.get('motion_types', [])
subtitle = first_nonempty(sample['annotation'].get('subtitle_segments'), task.get('subtitle_segments'), []) or []
rm = meta.get('reference_media') or {}
source_urls = urls(first_nonempty(source.get('source_urls'), rm.get('source_url'), seed.get('source_url')))
common = {
'sample_id': group_id, 'sequence_index': sample['source_number'],
'collection': sample['collection'], 'available_tasks': list(sample['payloads']),
'is_paired': len(sample['payloads']) > 1, 'summary': summary,
'video_category_ids': categories, 'video_categories': [tax['FAMILY_NAMES'].get(c, c) for c in categories],
'subtype_ids': subtypes, 'subtypes': [tax['SUBTYPE_NAMES'].get(s, s).rsplit(' / ', 1)[-1] for s in subtypes],
'subjects': as_strings(subjects), 'events': as_strings(events), 'motion_types': as_strings(motion),
'subject_label_count': len(subjects) if isinstance(subjects, list) else None,
'event_count': len(events) if isinstance(events, list) else None,
'construction_complexity': C, 'dynamic_complexity': D, 'dependency_span': S,
'source_urls': source_urls, 'source_platform': source.get('platform'),
'annotation_json': dumps(annotation), 'subtitle_segments_json': dumps(subtitle),
'difficulty_evidence_json': dumps({k: v for k, v in difficulty.items()
if k.endswith('_evidence') or k in ('assessment_source', 'evidence_scope')}),
}
result = []; mapping = []; issues = []
for setting in TASKS:
if setting not in sample['payloads']:
continue
payload = sample['payloads'][setting]
prompt = payload.get('prompt')
if not isinstance(prompt, str) or not prompt.strip():
raise ValueError(f'Missing final prompt: {sample["folder"]}/{setting}.json')
row = dict(common, task_id=f'{group_id}-{setting}', task=setting, prompt=prompt,
prompt_sha256=sha_text(prompt), reference_url=None, reference_description=None,
keyframe_timestamp_s=None, keyclip_start_s=None, keyclip_end_s=None,
review_reference_start_s=None, review_reference_end_s=None,
reference_time_basis=None, reference_kind=None, reference_locator_status='not_applicable',
reference_metadata_json=None,
task_events_json=dumps(first_nonempty(payload.get('task_events'), task.get('task_events'),
sample['annotation'].get('task_events'), events)),
task_attributes_json=dumps({k: v for k, v in payload.items() if k != 'prompt'}))
flags = []
if not summary:
flags.append('missing_summary')
if not categories or not subtypes:
flags.append('missing_taxonomy')
if not all((C, D, S)):
flags.append('missing_difficulty')
if not isinstance(events, list) or not events:
flags.append('missing_event_annotation')
if setting != 't2v':
recovered = (reference_sources or {}).get(row['task_id'], {})
row['reference_url'] = recovered.get('reference_url') or (source_urls[0] if source_urls else None)
row['reference_description'] = payload.get('reference_frame_summary' if setting == 'i2v' else 'reference_clip_summary')
row['reference_kind'] = 'image' if setting == 'i2v' else 'video_clip'
safe_rm = public_metadata(rm)
row['reference_metadata_json'] = dumps(safe_rm)
review_interval = rm.get('selected_interval_in_review_video_seconds') or {}
if isinstance(review_interval, dict):
row['review_reference_start_s'] = number(review_interval.get('start'))
row['review_reference_end_s'] = number(review_interval.get('end'))
if sample['collection'] == 'reproduce':
row['reference_time_basis'] = 'local_source_video_seconds'
if setting == 'i2v':
row['keyframe_timestamp_s'] = number(first_nonempty(payload.get('timestamp'),
(meta.get('generation') or {}).get('keyframe_timestamp')))
else:
row['keyclip_start_s'] = number(first_nonempty(payload.get('start_timestamp'),
(meta.get('generation') or {}).get('keyclip_start')))
row['keyclip_end_s'] = number(first_nonempty(payload.get('end_timestamp'),
(meta.get('generation') or {}).get('keyclip_end')))
else:
timestamps = first_nonempty(rm.get('source_timestamps_seconds'), seed.get('source_timestamp_seconds'))
row['reference_time_basis'] = 'original_image' if setting == 'i2v' else 'source_video_seconds'
if setting == 'i2v':
row['keyframe_timestamp_s'] = number(timestamps) if not isinstance(timestamps, (list, dict)) else None
else:
if isinstance(timestamps, dict):
row['keyclip_start_s'] = number(first_nonempty(timestamps.get('start'), timestamps.get('start_s')))
row['keyclip_end_s'] = number(first_nonempty(timestamps.get('end'), timestamps.get('end_s')))
elif isinstance(timestamps, list) and len(timestamps) == 2:
row['keyclip_start_s'], row['keyclip_end_s'] = map(number, timestamps)
# The local source used to construct the task is the original video.
# Preserve the actual selected interval recorded for that source.
if setting == 'v2v' and (row['keyclip_start_s'] is None or row['keyclip_end_s'] is None):
start, end = row.get('review_reference_start_s'), row.get('review_reference_end_s')
if start is not None and end is not None and 0 <= start < end:
row['keyclip_start_s'], row['keyclip_end_s'] = start, end
if setting == 'v2v' and recovered.get('keyclip_start_s') is not None and recovered.get('keyclip_end_s') is not None:
row['keyclip_start_s'] = number(recovered['keyclip_start_s'])
row['keyclip_end_s'] = number(recovered['keyclip_end_s'])
if not row['reference_description']:
flags.append('missing_reference_description')
if not row['reference_url']:
flags.append('missing_reference_url')
if setting == 'i2v' and sample['collection'] == 'reproduce' and row['keyframe_timestamp_s'] is None:
flags.append('missing_keyframe_timestamp')
if setting == 'v2v' and (row['keyclip_start_s'] is None or row['keyclip_end_s'] is None):
flags.append('missing_source_clip_timestamps')
times = [row[k] for k in ('keyframe_timestamp_s', 'keyclip_start_s', 'keyclip_end_s') if row[k] is not None]
if any(v < 0 for v in times):
flags.append('negative_reference_time')
if (row['keyclip_start_s'] is not None and row['keyclip_end_s'] is not None
and row['keyclip_start_s'] >= row['keyclip_end_s']):
flags.append('invalid_clip_interval')
row['reference_locator_status'] = 'missing_information' if any(
f in flags for f in ('missing_reference_url', 'missing_keyframe_timestamp', 'missing_source_clip_timestamps',
'negative_reference_time', 'invalid_clip_interval')) else 'recorded_not_url_verified'
row['data_warnings'] = flags
for flag in flags:
issues.append({'task_id': row['task_id'], 'sample_id': group_id, 'task': setting,
'collection': sample['collection'], 'issue': flag})
mapping.append({'task_id': row['task_id'], 'sample_id': group_id, 'sequence_index': sample['source_number'],
'original_sample_id': sample['original_id'], 'collection': sample['collection'], 'task': setting,
'input_folder': sample['folder'], 'prompt_path': str(Path(sample['folder']) / f'{setting}.json'),
'prompt_field': 'prompt', 'prompt_sha256': row['prompt_sha256'],
'annotation_path': str(Path(sample['folder']) / ('annotation.json' if sample['collection'] == 'reproduce' else 'task.json')),
'metadata_path': str(Path(sample['folder']) / 'metadata.json'),
'original_source_locator': first_nonempty(source.get('source_urls'), rm.get('source_url'), seed.get('source_url')),
'reference_media_path': str(Path(sample['folder']) / ('keyframe.png' if setting == 'i2v' else 'keyclip.mp4')) if setting != 't2v' else None})
result.append(row)
return result, mapping, issues
PUBLIC_COMMON_FIELDS = (
'sample_id', 'task_id', 'task', 'prompt', 'global_description',
'video_categories', 'subtypes', 'subjects', 'motion_types', 'events',
'construction_complexity', 'dynamic_complexity', 'dependency_span',
)
PUBLIC_REFERENCE_FIELDS = {
't2v': (),
'i2v': ('reference_url', 'reference_frame_summary', 'keyframe_timestamp_s'),
'v2v': ('reference_url', 'reference_clip_summary', 'keyclip_start_s', 'keyclip_end_s'),
}
PUBLIC_LIST_FIELDS = ('video_categories', 'subtypes', 'subjects', 'motion_types', 'events')
PUBLIC_OMITTED_FIELDS = frozenset({
'prompt_sha256', 'source_platform', 'data_warnings', 'sequence_index',
'subject_label_count', 'event_count', 'available_tasks', 'subtype_ids', 'video_category_ids',
'reference_locator_status', 'reference_metadata_json', 'difficulty_evidence_json',
'annotation_json', 'subtitle_segments_json', 'task_events_json', 'task_attributes_json',
'source_urls', 'collection', 'is_paired', 'review_reference_start_s', 'review_reference_end_s',
'reference_time_basis',
})
def public_record(row):
"""Export complete original annotation lists and task-specific reference information."""
result = dict(row)
result['global_description'] = row['summary']
if row['task'] == 'i2v':
result['reference_frame_summary'] = row.get('reference_description')
elif row['task'] == 'v2v':
result['reference_clip_summary'] = row.get('reference_description')
if result.get('keyclip_start_s') is None or result.get('keyclip_end_s') is None:
start, end = row.get('review_reference_start_s'), row.get('review_reference_end_s')
if start is not None and end is not None and 0 <= start < end:
result['keyclip_start_s'], result['keyclip_end_s'] = start, end
for key in PUBLIC_LIST_FIELDS:
value = result[key]
if not isinstance(value, list) or not all(isinstance(v, str) for v in value):
raise ValueError(f'{key} must be a list of original string annotations')
result[key] = list(value)
fields = PUBLIC_COMMON_FIELDS + PUBLIC_REFERENCE_FIELDS[row['task']]
published = {key: result.get(key) for key in fields}
published['qa'] = row['qa']
return published
def schema(setting=None):
import pyarrow as pa
fields = list(PUBLIC_COMMON_FIELDS)
for task in (TASKS if setting is None else (setting,)):
fields.extend(k for k in PUBLIC_REFERENCE_FIELDS[task] if k not in fields)
numeric = {'keyframe_timestamp_s', 'keyclip_start_s', 'keyclip_end_s'}
qa_type = pa.list_(pa.struct([(key, pa.string()) for key in
('section', 'dimension', 'name', 'question', 'expected_answer')]))
return pa.schema([(key, pa.list_(pa.string()) if key in PUBLIC_LIST_FIELDS else
pa.float64() if key in numeric else pa.string()) for key in fields]
+ [('qa', qa_type)])
QA_SPECIFIC = {'t2v': 'FullVideoSpecific', 'i2v': 'KeyframeSpecific', 'v2v': 'ReferenceClipSpecific'}
def flatten_task_qa(qa, setting):
"""Preserve saved QA order and exact text; select only this task's specific section."""
result = []
shared = qa.get('SharedChecklist')
if not isinstance(shared, dict):
raise ValueError('QA is missing SharedChecklist')
for group, dimensions in shared.items():
if not isinstance(dimensions, dict):
raise ValueError(f'Invalid QA group: {group}')
for dimension, item in dimensions.items():
if item is None:
continue
if not isinstance(item, dict) or not isinstance(item.get('question'), str) or not item['question'].strip():
raise ValueError(f'Invalid shared QA item: {group}.{dimension}')
if item.get('expected_answer') not in ('Yes', 'No'):
raise ValueError(f'Invalid QA expected answer: {group}.{dimension}')
result.append({'section': 'SharedChecklist', 'dimension': group + '.' + dimension,
'name': item.get('name'), 'question': item['question'],
'expected_answer': item['expected_answer']})
section = QA_SPECIFIC[setting]
specific = qa.get(section)
if not isinstance(specific, list) or not specific:
raise ValueError(f'QA is missing {section}')
for item in specific:
if not isinstance(item, dict) or not isinstance(item.get('question'), str) or not item['question'].strip():
raise ValueError(f'Invalid specific QA item: {section}')
if item.get('expected_answer') not in ('Yes', 'No'):
raise ValueError(f'Invalid QA expected answer: {section}')
result.append({'section': section, 'dimension': None, 'name': item.get('name'),
'question': item['question'], 'expected_answer': item['expected_answer']})
return result
def attach_task_qa(rows, mappings, input_root):
"""Use the same authoritative QA roots as the 9029 review page; never generate QA."""
index = {mapping['task_id']: mapping for mapping in mappings}
cache = {}
audit = []
for row in rows:
mapping = index[row['task_id']]
qa_root = 'code2video-qa-reproduce' if mapping['collection'] == 'reproduce' else 'code2video-qa-gen'
source = input_root / qa_root / mapping['original_sample_id'] / 'qa.json'
if source not in cache:
content = source.read_bytes()
cache[source] = (json.loads(content), hashlib.sha256(content).hexdigest())
qa, source_hash = cache[source]
row['qa'] = flatten_task_qa(qa, row['task'])
audit.append({'task_id': row['task_id'], 'source_path': str(source),
'source_sha256': source_hash, 'questions': len(row['qa']),
'specific_section': QA_SPECIFIC[row['task']]})
return audit
def validate_rows(rows):
seen = set(); pairs = {}
for row in rows:
assert row['task_id'] not in seen, row['task_id']
seen.add(row['task_id'])
assert sha_text(row['prompt']) == row['prompt_sha256']
assert '/mnt/disk/' not in dumps(row), row['task_id']
assert 'data:image/' not in dumps(row) and 'data:video/' not in dumps(row)
assert row['task_id'] == f'{row["sample_id"]}-{row["task"]}'
pairs.setdefault(row['sample_id'], []).append(row)
for sample_id, group in pairs.items():
if group[0]['collection'] == 'reproduce':
assert {r['task'] for r in group} == set(TASKS), sample_id
for name in ('summary', 'subjects', 'events', 'subtype_ids', 'construction_complexity',
'dynamic_complexity', 'dependency_span', 'sequence_index'):
assert len({dumps(r[name]) for r in group}) == 1, (sample_id, name)
def write_tables(public, rows, split=None):
import pyarrow as pa
import pyarrow.parquet as pq
files = []
for setting in TASKS:
subset = [public_record(r) for r in rows if r['task'] == setting]
folder = public / (split or 'all') / setting
folder.mkdir(parents=True, exist_ok=True)
parquet = folder / ('train-00000-of-00001.parquet' if split == 'train' else 'data-00000-of-00001.parquet')
table = pa.Table.from_pylist(subset, schema=schema(setting))
temporary = parquet.with_suffix('.parquet.tmp')
pq.write_table(table, temporary, compression='zstd')
restored = pq.read_table(temporary).to_pylist()
assert restored == table.to_pylist(), f'Parquet roundtrip mismatch: {parquet}'
temporary.replace(parquet)
jsonl = folder / 'records.jsonl'
jsonl.write_text(''.join(dumps(r) + '\n' for r in subset), encoding='utf-8')
write_csv(folder / 'records.csv', subset)
files.extend([parquet, jsonl, folder / 'records.csv'])
return files
def select_training_rows(rows, mappings, selected):
"""Require every manifest entry to resolve, and keep real pairs together."""
index = {m['task_id']: m for m in mappings}
criteria = {(r.get('collection') or None, r['sample_id'], r.get('task') or None) for r in selected}
matched = set(); chosen = set()
for row in rows:
m = index[row['task_id']]
for criterion in criteria:
collection, sample_id, setting = criterion
if (collection in (None, m['collection']) and sample_id == m['original_sample_id']
and setting in (None, m['task'])):
matched.add(criterion); chosen.add(row['task_id'])
if criteria - matched:
raise ValueError(f'Unresolved train manifest entries: {sorted(criteria-matched, key=str)}')
if not chosen:
raise ValueError('Explicit train manifest selects no tasks')
paired = {r['sample_id'] for r in rows if r['task_id'] in chosen and r['collection'] == 'reproduce'}
return [r for r in rows if r['task_id'] in chosen or r['sample_id'] in paired]
def public_readme(rows, train_mode):
template = Path(__file__).resolve().parents[1] / 'docs/README.template.md'
text = template.read_text(encoding='utf-8')
counts = Counter((r['collection'], r['task']) for r in rows)
by_task = Counter(r['task'] for r in rows)
sample = next((r for r in rows if r['task'] == 't2v' and 'Lody' in r['summary']),
next(r for r in rows if r['task'] == 't2v'))
siblings = {r['task']: r for r in rows if r['sample_id'] == sample['sample_id']}
example = {k: sample[k] for k in ('sample_id', 'summary', 'video_categories', 'subtypes', 'subjects',
'events', 'construction_complexity', 'dynamic_complexity', 'dependency_span')}
example['global_description'] = example.pop('summary')
count_rows = []
for collection, label in [('reproduce', 'Source-paired'), ('generate', 'Independent')]:
values = [counts[collection, t] for t in TASKS]
count_rows.append('| ' + label + ' | ' + ' | '.join(map(str, values)) + ' | ' + str(sum(values)) + ' |')
count_rows.append('| **Total** | ' + ' | '.join('**' + str(by_task[t]) + '**' for t in TASKS)
+ ' | **' + format(len(rows), ',') + '** |')
conditional = [r for r in rows if r['task'] != 't2v']
missing_urls = Counter((r['collection'], r['task']) for r in conditional if 'missing_reference_url' in r['data_warnings'])
missing_times = sum('missing_source_clip_timestamps' in r['data_warnings'] for r in rows)
notes = (f"All {len(rows):,} tasks have their existing summaries. "
f"Reference descriptions are present for {sum(bool(r['reference_description']) for r in conditional):,} "
f"of {len(conditional):,} conditional tasks. "
f"{sum(missing_urls.values())} conditional tasks lack recorded public reference URLs: "
f"{missing_urls['generate', 'i2v']} independent I2V, "
f"{missing_urls['generate', 'v2v']} independent V2V, and "
f"{missing_urls['reproduce', 'i2v'] + missing_urls['reproduce', 'v2v']} source-paired tasks. "
f"{missing_times} V2V tasks lack recorded original-source clip timestamps.")
def timestamp(setting, key):
value = siblings.get(setting, {}).get(key)
return f'{value:.3f}' if value is not None else 'not recorded'
replacements = {
'TOTAL_TASKS': format(len(rows), ','), 'FAMILY_COUNT': str(len({c for r in rows for c in r['video_category_ids']})),
'SUBTYPE_COUNT': str(len({s for r in rows for s in r['subtype_ids']})),
'PAIRED_GROUPS': str(len({r['sample_id'] for r in rows if r['collection'] == 'reproduce'})),
'INDEPENDENT_TASKS': str(sum(r['collection'] == 'generate' for r in rows)),
'COUNT_ROWS': '\n'.join(count_rows), 'EXAMPLE_SAMPLE_ID': sample['sample_id'],
'EXAMPLE_METADATA': json.dumps(example, ensure_ascii=False, indent=2),
'EXAMPLE_PROMPT': sample['prompt'], 'EXAMPLE_FRAME_TIME': timestamp('i2v', 'keyframe_timestamp_s'),
'EXAMPLE_CLIP_START': timestamp('v2v', 'keyclip_start_s'), 'EXAMPLE_CLIP_END': timestamp('v2v', 'keyclip_end_s'),
'COMPLETENESS_NOTES': notes,
**{t.upper() + '_COUNT': str(by_task[t]) for t in TASKS},
}
for key, value in replacements.items():
text = text.replace('{{' + key + '}}', value)
if re.search(r'\{\{[A-Z_]+\}\}', text):
raise ValueError('Unfilled dataset-card marker')
return text
def write_project_readmes(out, rows, train_mode):
public = out / 'huggingface'
card = public_readme(rows, train_mode)
(public / 'README.md').write_text(card, encoding='utf-8')
homepage = card.split('\n---\n', 1)[1].lstrip()
homepage = homepage.replace('src="assets/', 'src="huggingface/assets/')
for name in ('taxonomy.json', 'pair_index.csv', 'data_issues.csv', 'manifest.json', 'schema.json'):
homepage = homepage.replace('](' + name + ')', '](huggingface/' + name + ')')
homepage += ('\n---\n\nLocal archive and source-index documentation: '
'[LOCAL_EXPORT_README.md](LOCAL_EXPORT_README.md)。'
'Publish the `huggingface/` directory; keep `private/` and the private index archive local.\n')
(out / 'README.md').write_text(homepage, encoding='utf-8')
docs = public / 'docs'; docs.mkdir(exist_ok=True)
source = Path(__file__).resolve().parents[1] / 'docs/README.template.md'
(docs / 'README.template.md').write_text(source.read_text(), encoding='utf-8')
def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument('--project-root', type=Path, required=True)
parser.add_argument('--output-root', type=Path, default=Path(__file__).resolve().parents[1])
parser.add_argument('--workers', type=int, default=12)
parser.add_argument('--reuse-snapshot', action='store_true', help='Re-export a previously saved private input snapshot')
group = parser.add_mutually_exclusive_group()
group.add_argument('--hf-train-full', action='store_true', help='Explicitly use all rows as the HF train loading entry')
group.add_argument('--train-manifest', type=Path, help='Explicit training subset CSV: original sample_id, optional task and collection')
args = parser.parse_args()
project = args.project_root.resolve(); out = args.output_root.resolve()
public = out / 'huggingface'; private = out / 'private'
public.mkdir(parents=True, exist_ok=True); private.mkdir(parents=True, exist_ok=True)
snapshot = private / 'input_snapshot.jsonl'
if args.reuse_snapshot:
with snapshot.open() as stream:
samples = [json.loads(line) for line in stream]
else:
jobs = list_samples(project / 'code2video-input')
samples = []
print(f'Reading {len(jobs)} sample groups; metadata/text only; no media/API calls', flush=True)
with ThreadPoolExecutor(max_workers=args.workers) as pool:
for i, sample in enumerate(pool.map(collect_sample, jobs), 1):
samples.append(sample)
if i % 50 == 0 or i == len(jobs):
print(f'Inputs {i}/{len(jobs)}', flush=True)
snapshot.write_text(''.join(dumps(s) + '\n' for s in samples), encoding='utf-8')
tax = taxonomy(project); manifest = source_manifest(project)
reference_sources = read_json(public / 'reference_sources.json')
rows = []; mappings = []; issues = []
for sample in samples:
r, m, flags = make_records(sample, manifest, tax, reference_sources)
rows.extend(r); mappings.extend(m); issues.extend(flags)
rows.sort(key=lambda r: (r['sequence_index'], TASKS.index(r['task'])))
qa_audit = attach_task_qa(rows, mappings, project / 'code2video-input')
write_json(private / 'qa_export_audit.json', qa_audit)
validate_rows(rows)
counts = Counter(r['task'] for r in rows)
print(f'Validated {len(rows)} tasks: {dict(counts)}', flush=True)
files = write_tables(public, rows)
train_mode = 'not yet specified; no independent training subset inferred'
if args.hf_train_full:
files.extend(write_tables(public, rows, 'train'))
train_mode = 'HF train contains the same complete records as all; not a held-out split'
elif args.train_manifest:
selected = read_csv(args.train_manifest)
train_rows = select_training_rows(rows, mappings, selected)
files.extend(write_tables(public, train_rows, 'train'))
train_mode = f'explicit manifest; {len(train_rows)} rows; reproduction pairing kept together'
write_csv(private / 'source_index.csv', mappings)
(private / 'source_index.jsonl').write_text(''.join(dumps(m) + '\n' for m in mappings), encoding='utf-8')
write_csv(public / 'data_issues.csv', issues)
write_json(public / 'taxonomy.json', tax)
write_json(public / 'schema.json', {task: {f.name: str(f.type) for f in schema(task)} for task in TASKS})
write_project_readmes(out, rows, train_mode)
info = {'schema_version': VERSION, 'built_at': datetime.now(timezone.utc).isoformat(),
'sample_groups': len(samples), 'task_count': len(rows), 'by_task': dict(counts),
'by_collection_and_task': dict(Counter(r['collection'] + '/' + r['task'] for r in rows)),
'train_mode': train_mode, 'warnings': dict(Counter(i['issue'] for i in issues)),
'rows_with_warnings': sum(bool(r['data_warnings']) for r in rows),
'prompt_fields': ['t2v.json:prompt', 'i2v.json:prompt', 'v2v.json:prompt'],
'media_embedded': False, 'shuffled': False, 'api_calls': 0,
'source_input_index_sha256': hashlib.sha256((project/'code2video-input/index.csv').read_bytes()).hexdigest(),
'pyarrow_version': __import__('pyarrow').__version__,
'list_presentation': 'Native lists of complete original string entries; no count replacement',
'public_omitted_fields': sorted(PUBLIC_OMITTED_FIELDS),
'documentation_asset_source': 'code2video-compass-release.zip',
'documentation_files': {str(f.relative_to(public)): {'bytes': f.stat().st_size,
'sha256': hashlib.sha256(f.read_bytes()).hexdigest()} for f in
[public/'README.md', public/'docs/README.template.md', *sorted((public/'assets').glob('*'))] if f.is_file()},
'files': {str(f.relative_to(public)): {'bytes': f.stat().st_size,
'sha256': hashlib.sha256(f.read_bytes()).hexdigest()} for f in files}}
write_json(public / 'manifest.json', info)
paired = []
for s in samples:
sample_id = f'C2VC-{s["source_number"]:04d}'
paired.append({'sample_id': sample_id, 'collection': s['collection'],
**{t: f'{sample_id}-{t}' if t in s['payloads'] else None for t in TASKS}})
write_csv(public / 'pair_index.csv', paired)
(out / 'LOCAL_EXPORT_README.md').write_text(f'''# Code2Video-Compass local release bundle
## Contents
- {len(samples)} source or independent task groups and {len(rows)} tasks: T2V {counts['t2v']}, I2V {counts['i2v']}, V2V {counts['v2v']}.
- `huggingface/`: public files, with separate Parquet, JSONL, and CSV tables for each setting, descriptions, category labels, subject types, events, and C/D/S annotations.
- `huggingface/pair_index.csv`: public IDs for corresponding source-paired tasks. The original order is preserved. Independent tasks do not receive invented counterparts.
- `private/source_index.csv`: public ID to original sample ID, input folder, prompt field, and local reference path. Keep this file private.
- `private/input_snapshot.jsonl`: a read-only snapshot of source instructions and metadata. Keep this file private.
- T2V keeps its task prompt; I2V/V2V also contain reference URLs, recorded timestamps, and reference descriptions.
- Source data is preserved. Media is not embedded or copied, models are not called, and nothing is uploaded by this builder. Each task includes its applicable QA; common code-generation templates are excluded.
- Descriptions reuse existing annotations. Subjects keep the original `primary_subject` type labels rather than entity counts.
- Missing information remains null. See `huggingface/data_issues.csv` and `huggingface/manifest.json`.
## Train and all
{train_mode}
The existing half-sample CSV files define evaluation scopes, not training sets.
`all/` contains the complete tasks; `train` in the HF card is a loading split.
Supply an explicit manifest for a separate training subset, keeping the three
settings of each source-paired group together.
## Rebuild
```bash
cd {out}
python scripts/build_release.py \\
--project-root {project} --workers 12 --hf-train-full
# For a separate training subset, replace --hf-train-full with --train-manifest /path/to/train.csv.
# Add --reuse-snapshot to rebuild from the saved input snapshot.
```
The organization of HF configs, documentation, and field descriptions follows
the supplied ChartAnno example archives. Project names, licenses, figures,
citations, and data from those examples are not copied into this release.
''', encoding='utf-8')
script_dest = public / 'scripts'
script_dest.mkdir(exist_ok=True)
(script_dest / 'build_release.py').write_text(Path(__file__).read_text(), encoding='utf-8')
(public / 'requirements.txt').write_text(f'pyarrow=={__import__("pyarrow").__version__}\n', encoding='utf-8')
archive = out / 'code2video_compass_hf.zip'
temporary_archive = archive.with_suffix('.zip.tmp')
with zipfile.ZipFile(temporary_archive, 'w', compression=zipfile.ZIP_DEFLATED) as zipped:
for p in sorted(public.rglob('*')):
if p.is_file() and not p.name.endswith('.tmp'):
assert p.suffix not in ('.mp4', '.webm')
assert p.suffix not in ('.png', '.jpg', '.svg') or p.relative_to(public).parts[0] == 'assets'
zipped.write(p, 'code2video_compass/' + p.relative_to(public).as_posix())
temporary_archive.replace(archive)
print(json.dumps({'release': str(public), 'archive': str(archive), **info}, ensure_ascii=False, indent=2), flush=True)
if __name__ == '__main__':
main()
|