Image-Text-to-Text
Transformers
Safetensors
qwen3_5
vllm
video
multimodal
reinforcement-learning
temporal-grounding
object-tracking
video-segmentation
visual-question-answering
spatial-reasoning
qwen3.5
conversational
Instructions to use OraRL/Video-ORA-4B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use OraRL/Video-ORA-4B with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("image-text-to-text", model="OraRL/Video-ORA-4B") messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] pipe(text=messages)# Load model directly from transformers import AutoProcessor, AutoModelForMultimodalLM processor = AutoProcessor.from_pretrained("OraRL/Video-ORA-4B") model = AutoModelForMultimodalLM.from_pretrained("OraRL/Video-ORA-4B", device_map="auto") messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] inputs = processor.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt", ).to(model.device) outputs = model.generate(**inputs, max_new_tokens=40) print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:])) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use OraRL/Video-ORA-4B with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "OraRL/Video-ORA-4B" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "OraRL/Video-ORA-4B", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker
docker model run hf.co/OraRL/Video-ORA-4B
- SGLang
How to use OraRL/Video-ORA-4B with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "OraRL/Video-ORA-4B" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "OraRL/Video-ORA-4B", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "OraRL/Video-ORA-4B" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "OraRL/Video-ORA-4B", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }' - Docker Model Runner
How to use OraRL/Video-ORA-4B with Docker Model Runner:
docker model run hf.co/OraRL/Video-ORA-4B
File size: 60,841 Bytes
0185029 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 | """Benchmark adapters that produce portable canonical evaluation rows."""
from __future__ import annotations
import ast
import base64
import binascii
import csv
import hashlib
import io
import json
import math
import os
import re
import sys
from collections.abc import Mapping, Sequence
from concurrent.futures import ThreadPoolExecutor
from dataclasses import dataclass
from pathlib import Path
from typing import Any
from orarl.data.schema import MEDIA_PATH_KEYS, read_records
from .layout import annotation_path, artifact_directory, media_directory, sha256_file
from .schema import EVALUATION_SCHEMA_VERSION, EvaluationSchemaError, validate_evaluation_rows
from .sources import GENERIC_ADAPTERS, EvaluationSource
_PATH_KEY_RE = re.compile(
r"(?:^|_)(?:artifact|artifacts|dir|directory|file|files|image|images|media|"
r"path|paths|root|subtitle|subtitles|tensor|tensors|video|videos)$"
)
_WINDOWS_ABSOLUTE_RE = re.compile(r"^[A-Za-z]:[\\/]")
_IMAGE_SUFFIXES = frozenset({".bmp", ".gif", ".jpeg", ".jpg", ".png", ".tif", ".webp"})
_VIDEO_SUFFIXES = frozenset({".avi", ".mkv", ".mov", ".mp4", ".mpeg", ".webm"})
_SUBTITLE_SUFFIXES = (".srt", ".vtt", ".json", ".txt")
_ARTIFACT_SUFFIXES = (".npz", ".npy", ".safetensors", ".pt", ".pth", ".pkl")
_GROUND_TRUTH_ANSWER_ALIAS = "_".join(("gt", "answer"))
_GROUND_TRUTH_BBOX_ALIAS = "_".join(("gt", "bbox"))
_MINDCUBE_EXPECTED_GROUP_COUNTS = {
"rotation": 200,
"among": 600,
"around": 250,
}
_REVSI_EXPECTED_GROUP_COUNTS = {
"object_abs_distance": 1232,
"object_counting": 1091,
"object_rel_direction": 1387,
"object_rel_distance": 1351,
"object_size_estimation": 1134,
"room_size_estimation": 302,
"route_planning": 311,
}
_PAYLOAD_KEYS = frozenset(
{
"bbox",
"bboxes",
"boxes",
"coordinates",
"duration",
"end",
"frame",
"frame_id",
"frame_ids",
_GROUND_TRUTH_BBOX_ALIAS,
"mask",
"masks",
"normalized_solution",
"object",
"segment",
"segmentation",
"segmentation_output",
"span",
"spans",
"start",
"target",
"timestamp",
"timestamps",
"trajectory",
}
)
_BASE_CONSUMED_FIELDS = frozenset(
{
"answer",
"answers",
"bench",
"benchmark",
"caption",
"choices",
"data_source",
"data_type",
"dataset",
"eval_task",
"evaluation",
"family",
"file_name",
"ground_truth",
_GROUND_TRUTH_ANSWER_ALIAS,
_GROUND_TRUTH_BBOX_ALIAS,
"id",
"image",
"image_path",
"images",
"index",
"input_prompt",
"media",
"media_path",
"messages",
"metadata",
"normal_caption",
"options",
"oracle_label",
"oracle_labels",
"path",
"preprocessed",
"preprocessed_video",
"problem",
"problem_id",
"problem_type",
"prompt",
"question",
"question_id",
"question_type",
"sample_id",
"schema_version",
"solution",
"source",
"split",
"subtitle_file",
"subtitle_path",
"subtitles",
"task_payload",
"task_type",
"video",
"video_path",
"videos",
}
)
_DROP = object()
class ConversionError(ValueError):
"""Raised when a benchmark source cannot become canonical rows."""
@dataclass(frozen=True)
class PlannedAsset:
"""One content-addressed output asset and its private local source."""
repository_path: str
sha256: str | None
bytes: int
kind: str
benchmark: str
license: str
source_path: Path | None = None
content: bytes | None = None
missing: bool = False
@dataclass(frozen=True)
class ConvertedSource:
"""Canonical rows and assets produced for one source record."""
source: EvaluationSource
rows: tuple[dict[str, Any], ...]
assets: tuple[PlannedAsset, ...]
source_rows: int
@property
def missing_assets(self) -> tuple[PlannedAsset, ...]:
return tuple(asset for asset in self.assets if asset.missing)
def _json_bytes(value: Any) -> bytes:
return json.dumps(
value,
ensure_ascii=False,
sort_keys=True,
separators=(",", ":"),
allow_nan=False,
).encode("utf-8")
def _canonical_suffix(path: Path) -> str:
suffix = path.suffix.casefold()
aliases = {".jpeg": ".jpg", ".tiff": ".tif"}
suffix = aliases.get(suffix, suffix)
if not re.fullmatch(r"\.[a-z0-9]{1,12}", suffix):
return ""
return suffix
def _image_suffix(content: bytes) -> str:
if content.startswith(b"\xff\xd8\xff"):
return ".jpg"
if content.startswith(b"\x89PNG\r\n\x1a\n"):
return ".png"
if content.startswith((b"GIF87a", b"GIF89a")):
return ".gif"
if content.startswith(b"BM"):
return ".bmp"
if content[:4] in {b"II*\x00", b"MM\x00*"}:
return ".tif"
if content.startswith(b"RIFF") and content[8:12] == b"WEBP":
return ".webp"
return ".bin"
def _is_path_key(key: object) -> bool:
return isinstance(key, str) and bool(_PATH_KEY_RE.search(key.casefold()))
def _is_private_path(value: str, source: EvaluationSource) -> bool:
stripped = value.strip()
if (
stripped.startswith(("/", "~/", "file://"))
or _WINDOWS_ABSOLUTE_RE.match(stripped)
):
return True
private_roots = {
str(source.annotation_input),
str(source.annotation_input.parent),
*(str(path) for path in source.media_roots.values()),
*(str(path) for path in source.preprocessed_roots.values()),
}
return any(root and root in stripped for root in private_roots)
def _json_safe(value: Any, source: EvaluationSource) -> Any:
if value is None or isinstance(value, (bool, int)):
return value
if isinstance(value, float):
return value if math.isfinite(value) else _DROP
if isinstance(value, str):
return _DROP if _is_private_path(value, source) else value
if isinstance(value, (bytes, bytearray, memoryview, Path)):
return _DROP
if isinstance(value, Mapping):
converted: dict[str, Any] = {}
for raw_key, item in value.items():
if not isinstance(raw_key, str) or _is_private_path(raw_key, source):
continue
safe = _json_safe(item, source)
if safe is not _DROP:
converted[raw_key] = safe
return converted
if isinstance(value, Sequence):
converted_items = []
for item in value:
safe = _json_safe(item, source)
if safe is not _DROP:
converted_items.append(safe)
return converted_items
item_method = getattr(value, "item", None)
if callable(item_method):
try:
scalar = item_method()
except (TypeError, ValueError):
scalar = value
if scalar is not value:
return _json_safe(scalar, source)
list_method = getattr(value, "tolist", None)
if callable(list_method):
try:
listed = list_method()
except (TypeError, ValueError):
listed = value
if listed is not value:
return _json_safe(listed, source)
isoformat = getattr(value, "isoformat", None)
if callable(isoformat):
try:
return str(isoformat())
except (TypeError, ValueError):
pass
return _DROP
def _as_list(value: Any) -> list[Any]:
if value is None:
return []
list_method = getattr(value, "tolist", None)
if callable(list_method) and not isinstance(value, (str, bytes, bytearray)):
try:
value = list_method()
except (TypeError, ValueError):
pass
if isinstance(value, Sequence) and not isinstance(value, (str, bytes, bytearray)):
return list(value)
return [value]
def _is_blank_path_reference(value: Any) -> bool:
if not isinstance(value, (str, os.PathLike)):
return False
raw = os.fspath(value)
return not raw.strip()
def _first(record: Mapping[str, Any], fields: Sequence[str]) -> Any:
for field in fields:
value = record.get(field)
if value is not None and not _is_blank_path_reference(value):
return value
return None
def _message_content(
record: Mapping[str, Any],
role: str,
*,
reverse: bool = False,
) -> Any:
messages = record.get("messages")
if not isinstance(messages, list):
return None
candidates = reversed(messages) if reverse else messages
for message in candidates:
if not isinstance(message, Mapping):
continue
if str(message.get("role") or "").casefold() != role:
continue
content = message.get("content")
if isinstance(content, str):
return content
if isinstance(content, list):
text = "\n".join(
str(item.get("text"))
for item in content
if isinstance(item, Mapping) and item.get("text") is not None
).strip()
if text:
return text
return None
class _AssetPlanner:
def __init__(
self,
source: EvaluationSource,
*,
hash_assets: bool = False,
):
self.source = source
self.hash_assets = hash_assets
self._assets: dict[str, PlannedAsset] = {}
self._source_paths: dict[tuple[str, str], str] = {}
self._missing_paths: dict[tuple[str, str], str] = {}
self._references: dict[tuple[str, str, str], tuple[str, Path]] = {}
self._resolved_paths: dict[tuple[str, str], Path] = {}
self._embedded_counts: dict[str, int] = {}
@property
def assets(self) -> tuple[PlannedAsset, ...]:
return tuple(
sorted(
self._assets.values(),
key=lambda asset: (asset.repository_path, asset.missing),
)
)
def _destination(self, kind: str, digest: str, suffix: str) -> str:
filename = digest + suffix
if kind == "artifacts":
return f"{artifact_directory(self.source.benchmark)}/{filename}"
return f"{media_directory(self.source.benchmark, kind)}/{filename}"
def _source_destination(self, kind: str, path: Path, root: Path) -> str:
try:
relative = path.relative_to(root.resolve())
except ValueError:
relative = Path(path.name)
raw_name = "__".join(relative.parts)
filename = re.sub(r"[^A-Za-z0-9._+-]+", "_", raw_name).strip("_")
if not filename or filename in {".", ".."}:
raise ConversionError(f"asset path has no portable filename: {path}")
if len(filename.encode("utf-8")) > 220:
suffix = _canonical_suffix(path)
stem = filename[: -len(suffix)] if suffix and filename.endswith(suffix) else filename
path_token = hashlib.blake2b(
relative.as_posix().encode("utf-8"),
digest_size=10,
).hexdigest()
stem_limit = 220 - len(suffix.encode("utf-8")) - len(path_token) - 2
filename = f"{stem[:stem_limit]}--{path_token}{suffix}"
if kind == "artifacts":
directory = artifact_directory(self.source.benchmark)
else:
directory = media_directory(self.source.benchmark, kind)
return f"{directory}/{self.source.split}/{filename}"
@staticmethod
def _reference_key(
reference: object,
root: Path,
kind: str,
) -> tuple[str, str, str]:
if not isinstance(reference, (str, os.PathLike)):
raise ConversionError("asset reference must be a local path string")
raw_reference = os.fspath(reference).strip()
return (
kind,
os.path.normcase(os.path.abspath(os.path.normpath(str(root)))),
os.path.normcase(os.path.normpath(raw_reference)),
)
def resolve(self, reference: object, root: Path) -> Path:
if not isinstance(reference, (str, os.PathLike)):
raise ConversionError("asset reference must be a local path string")
raw = os.fspath(reference).strip()
if not raw:
raise ConversionError("asset reference must be nonempty")
if re.match(r"^[A-Za-z][A-Za-z0-9+.-]*://", raw):
raise ConversionError(f"remote asset references are unsupported: {raw}")
resolved_key = (
os.path.normcase(os.path.abspath(os.path.normpath(str(root)))),
os.path.normcase(os.path.normpath(raw)),
)
cached = self._resolved_paths.get(resolved_key)
if cached is not None:
return cached
candidate = Path(os.path.expanduser(raw))
if candidate.is_absolute():
candidates = [candidate]
else:
normalized = Path(os.path.normpath(raw))
candidates = [root / normalized]
if normalized.parts and normalized.parts[0].casefold() == root.name.casefold():
candidates.append(root.parent / normalized)
if normalized.parts and normalized.parts[0].casefold() == "evaluation":
candidates.append(root.joinpath(*normalized.parts[1:]))
candidates.extend(
(
self.source.annotation.parent / normalized,
root / normalized.name,
)
)
unique: list[Path] = []
seen: set[str] = set()
for path in candidates:
normalized_path = Path(os.path.abspath(os.path.normpath(str(path))))
key = os.path.normcase(str(normalized_path))
if key not in seen:
seen.add(key)
unique.append(normalized_path)
for path in unique:
if path.is_file():
self._resolved_paths[resolved_key] = path
return path
self._resolved_paths[resolved_key] = unique[0]
return unique[0]
def mapped_path(
self,
reference: object,
root: Path,
*,
kind: str,
) -> str | None:
reference_key = self._reference_key(reference, root, kind)
cached = self._references.get(reference_key)
if cached is not None:
return cached[0]
path = self.resolve(reference, root)
mapped = self._source_paths.get((kind, os.path.normcase(str(path))))
if mapped is not None:
self._references[reference_key] = (mapped, path)
return mapped
def add_file(self, reference: object, *, kind: str, root: Path) -> tuple[str, Path]:
reference_key = self._reference_key(reference, root, kind)
cached = self._references.get(reference_key)
if cached is not None:
return cached
path = self.resolve(reference, root)
source_key = os.path.normcase(str(path))
mapped = self._source_paths.get((kind, source_key))
if mapped is not None:
result = (mapped, path)
self._references[reference_key] = result
return result
if not path.is_file():
missing_key = (kind, source_key)
repository_path = self._missing_paths.get(missing_key)
if repository_path is None:
token = hashlib.sha256(f"{kind}\0{source_key}".encode("utf-8")).hexdigest()
repository_path = self._destination(
kind,
f"missing-{token}",
_canonical_suffix(path),
)
self._missing_paths[missing_key] = repository_path
self._assets[repository_path] = PlannedAsset(
repository_path=repository_path,
sha256=None,
bytes=0,
kind=kind,
benchmark=self.source.benchmark,
license=self.source.license,
source_path=path,
missing=True,
)
self._source_paths[(kind, source_key)] = repository_path
result = (repository_path, path)
self._references[reference_key] = result
return result
resolved = path.resolve(strict=True)
if not resolved.is_file():
raise ConversionError(f"asset source is not a regular file: {path}")
byte_count = resolved.stat().st_size
if self.hash_assets:
digest = sha256_file(resolved)
repository_path = self._destination(
kind,
digest,
_canonical_suffix(resolved),
)
else:
digest = None
repository_path = self._source_destination(kind, resolved, root)
asset = PlannedAsset(
repository_path=repository_path,
sha256=digest,
bytes=byte_count,
kind=kind,
benchmark=self.source.benchmark,
license=self.source.license,
source_path=resolved,
)
previous = self._assets.get(repository_path)
if previous is not None and (
previous.sha256 != asset.sha256
or previous.bytes != asset.bytes
or (
asset.sha256 is None
and previous.source_path != asset.source_path
)
):
raise ConversionError(f"asset output path collision: {repository_path}")
self._assets[repository_path] = asset
self._source_paths[(kind, source_key)] = repository_path
self._source_paths[
(kind, os.path.normcase(str(resolved)))
] = repository_path
result = (repository_path, resolved)
self._references[reference_key] = result
return result
def add_bytes(
self,
content: bytes,
*,
kind: str,
suffix: str = "",
) -> str:
if self.hash_assets:
digest = hashlib.sha256(content).hexdigest()
repository_path = self._destination(kind, digest, suffix)
else:
digest = None
index = self._embedded_counts.get(kind, 0) + 1
self._embedded_counts[kind] = index
filename = f"{self.source.split}__embedded_{index:08d}{suffix}"
if kind == "artifacts":
directory = artifact_directory(self.source.benchmark)
else:
directory = media_directory(self.source.benchmark, kind)
repository_path = f"{directory}/{self.source.split}/{filename}"
asset = PlannedAsset(
repository_path=repository_path,
sha256=digest,
bytes=len(content),
kind=kind,
benchmark=self.source.benchmark,
license=self.source.license,
content=content,
)
previous = self._assets.get(repository_path)
if previous is not None and (
previous.sha256 != asset.sha256 or previous.bytes != asset.bytes
):
raise ConversionError(f"asset output path collision: {repository_path}")
self._assets[repository_path] = asset
return repository_path
def _kind_for_key(key: str, default: str = "artifacts") -> str:
folded = key.casefold()
if "subtitle" in folded:
return "subtitles"
if "image" in folded:
return "images"
if "video" in folded:
return "videos"
return default
def _root_for_kind(source: EvaluationSource, kind: str, key: str = "") -> Path:
if kind == "artifacts":
return source.preprocessed_root(key)
return source.media_root(kind)
def _stage_path_value(
value: Any,
*,
key: str,
source: EvaluationSource,
planner: _AssetPlanner,
default_kind: str = "artifacts",
force_kind: str | None = None,
) -> Any:
kind = force_kind or _kind_for_key(key, default_kind)
root = _root_for_kind(source, kind, key)
if isinstance(value, (str, os.PathLike)):
if _is_blank_path_reference(value):
return _DROP
local_path = planner.resolve(value, root)
annotation_sources = {
os.path.normcase(str(source.annotation)),
os.path.normcase(str(source.annotation.resolve())),
}
if os.path.normcase(str(local_path)) in annotation_sources:
return annotation_path(source.benchmark, source.split)
mapped = planner.mapped_path(value, root, kind=kind)
if mapped is not None:
return mapped
repository_path, _local = planner.add_file(value, kind=kind, root=root)
return repository_path
if isinstance(value, Mapping):
staged: dict[str, Any] = {}
for raw_key, item in value.items():
if not isinstance(raw_key, str) or _is_private_path(raw_key, source):
continue
converted = _stage_path_value(
item,
key=raw_key,
source=source,
planner=planner,
default_kind=kind,
force_kind=force_kind,
)
if converted is not _DROP:
staged[raw_key] = converted
return staged
if isinstance(value, Sequence) and not isinstance(value, (str, bytes, bytearray)):
staged_items = []
for item in value:
converted = _stage_path_value(
item,
key=key,
source=source,
planner=planner,
default_kind=kind,
force_kind=force_kind,
)
if converted is not _DROP:
staged_items.append(converted)
return staged_items
raise ConversionError(f"{key} must contain a local path string or list")
def _stage_payload(
value: Any,
*,
source: EvaluationSource,
planner: _AssetPlanner,
path_kind: str | None = None,
) -> Any:
if isinstance(value, Mapping):
staged: dict[str, Any] = {}
for raw_key, item in value.items():
if not isinstance(raw_key, str) or _is_private_path(raw_key, source):
continue
if _is_path_key(raw_key):
converted = _stage_path_value(
item,
key=raw_key,
source=source,
planner=planner,
force_kind=path_kind,
)
if converted is not _DROP:
staged[raw_key] = converted
continue
nested = _stage_payload(
item,
source=source,
planner=planner,
path_kind=path_kind,
)
if nested is not _DROP:
staged[raw_key] = nested
return staged
if isinstance(value, Sequence) and not isinstance(value, (str, bytes, bytearray)):
staged_items = []
for item in value:
nested = _stage_payload(
item,
source=source,
planner=planner,
path_kind=path_kind,
)
if nested is not _DROP:
staged_items.append(nested)
return staged_items
return _json_safe(value, source)
def _profile_without_paths(value: Any, source: EvaluationSource) -> Any:
if isinstance(value, Mapping):
converted: dict[str, Any] = {}
for raw_key, item in value.items():
if (
not isinstance(raw_key, str)
or _is_private_path(raw_key, source)
or _is_path_key(raw_key)
):
continue
safe = _profile_without_paths(item, source)
if safe is not _DROP:
converted[raw_key] = safe
return converted
if isinstance(value, Sequence) and not isinstance(value, (str, bytes, bytearray)):
converted_items = []
for item in value:
safe = _profile_without_paths(item, source)
if safe is not _DROP:
converted_items.append(safe)
return converted_items
return _json_safe(value, source)
def _structured_answer_payload(value: Any) -> Mapping[str, Any] | None:
if isinstance(value, Mapping):
return value
if not isinstance(value, str):
return None
matches = re.findall(
r"<answer>\s*(.*?)\s*</answer>",
value,
flags=re.IGNORECASE | re.DOTALL,
)
candidate = matches[-1] if matches else value.strip()
if candidate.startswith("```"):
candidate = re.sub(r"^```(?:json)?\s*", "", candidate, flags=re.IGNORECASE)
candidate = re.sub(r"\s*```$", "", candidate)
try:
parsed = json.loads(candidate)
except json.JSONDecodeError:
return None
return parsed if isinstance(parsed, Mapping) else None
def _stage_media_entry(
entry: Any,
*,
kind: str,
source: EvaluationSource,
planner: _AssetPlanner,
) -> tuple[Any, Path]:
root = source.media_root(kind)
if isinstance(entry, (str, os.PathLike)):
if _is_blank_path_reference(entry):
raise ConversionError(f"{kind} entry contains an empty local path")
repository_path, local_path = planner.add_file(entry, kind=kind, root=root)
return repository_path, local_path
if not isinstance(entry, Mapping):
raise ConversionError(f"{kind} entry must be a path string or JSON object")
staged: dict[str, Any] = {}
local_path: Path | None = None
for raw_key, value in entry.items():
if not isinstance(raw_key, str) or _is_private_path(raw_key, source):
continue
if raw_key in MEDIA_PATH_KEYS and isinstance(value, (str, os.PathLike)):
if _is_blank_path_reference(value):
continue
repository_path, candidate = planner.add_file(value, kind=kind, root=root)
staged[raw_key] = repository_path
if local_path is None:
local_path = candidate
else:
safe = _json_safe(value, source)
if safe is not _DROP:
staged[raw_key] = safe
if local_path is None:
raise ConversionError(f"{kind} entry does not contain a local path")
return staged, local_path
def _media_inputs(record: Mapping[str, Any], source: EvaluationSource) -> tuple[Any, Any]:
images = _first(record, ("images", "image", "image_path"))
videos = _first(record, ("videos", "video", "video_path"))
generic = _first(record, ("path", "media_path", "media", "file_name"))
if source.eval_task == "mvbench" and generic is not None:
# MVBench keeps the upstream WebM filename in ``video`` while ``path``
# points to the materialized MP4 used by its evaluator.
videos = generic
if generic is not None and images is None and videos is None:
values = _as_list(generic)
suffixes = {
Path(str(value)).suffix.casefold()
for value in values
if isinstance(value, (str, os.PathLike))
}
data_type = str(record.get("data_type") or "").casefold()
family = source.family.casefold()
if suffixes and suffixes <= _IMAGE_SUFFIXES:
images = generic
elif (
"image" in data_type
or source.adapter == "spatial_grounding"
or ("spatial_grounding" in family and "temporal" not in family)
):
images = generic
else:
videos = generic
return images, videos
def _stage_media(
record: Mapping[str, Any],
source: EvaluationSource,
planner: _AssetPlanner,
) -> tuple[list[Any], list[Any], list[Path]]:
raw_images, raw_videos = _media_inputs(record, source)
images: list[Any] = []
videos: list[Any] = []
video_sources: list[Path] = []
for entry in _as_list(raw_images):
if _is_blank_path_reference(entry):
continue
staged, _local = _stage_media_entry(
entry,
kind="images",
source=source,
planner=planner,
)
images.append(staged)
for entry in _as_list(raw_videos):
if _is_blank_path_reference(entry):
continue
staged, local = _stage_media_entry(
entry,
kind="videos",
source=source,
planner=planner,
)
videos.append(staged)
video_sources.append(local)
return images, videos, video_sources
def _subtitle_input(record: Mapping[str, Any], source: EvaluationSource) -> tuple[Any, bool]:
explicit = _first(record, ("subtitles", "subtitle_path", "subtitle_file"))
if explicit is not None:
return explicit, True
subtitle = record.get("subtitle")
if subtitle is None:
return None, False
values = _as_list(subtitle)
if all(
isinstance(value, str)
and (
Path(value).suffix.casefold() in _SUBTITLE_SUFFIXES
or source.media_root("subtitles").joinpath(value).is_file()
)
for value in values
):
return subtitle, True
return None, False
def _derived_files(
source_paths: Sequence[Path],
root: Path,
suffixes: Sequence[str],
video_root: Path | None = None,
) -> list[Path]:
derived: list[Path] = []
seen: set[str] = set()
for source_path in source_paths:
relative = Path(source_path.name)
if video_root is not None:
try:
relative = source_path.relative_to(video_root)
except ValueError:
pass
found = False
for suffix in suffixes:
candidates = (
root / relative.with_suffix(suffix),
root / f"{source_path.stem}{suffix}",
)
for candidate in candidates:
if not candidate.is_file():
continue
key = os.path.normcase(str(candidate.resolve()))
if key not in seen:
seen.add(key)
derived.append(candidate)
found = True
break
if found:
break
return derived
def _stage_subtitles(
record: Mapping[str, Any],
source: EvaluationSource,
planner: _AssetPlanner,
video_sources: Sequence[Path],
) -> tuple[list[Any], bool]:
raw, consumed_subtitle = _subtitle_input(record, source)
subtitles: list[Any] = []
if raw is not None:
for entry in _as_list(raw):
if _is_blank_path_reference(entry):
continue
staged, _local = _stage_media_entry(
entry,
kind="subtitles",
source=source,
planner=planner,
)
subtitles.append(staged)
elif "subtitles" in source.media_roots:
for path in _derived_files(
video_sources,
source.media_roots["subtitles"],
_SUBTITLE_SUFFIXES,
source.media_roots.get("videos", source.media_roots.get("default")),
):
staged, _local = _stage_media_entry(
path,
kind="subtitles",
source=source,
planner=planner,
)
subtitles.append(staged)
return subtitles, consumed_subtitle
def _stage_preprocessed(
record: Mapping[str, Any],
source: EvaluationSource,
planner: _AssetPlanner,
video_sources: Sequence[Path],
) -> dict[str, Any]:
raw = record.get("preprocessed")
staged: dict[str, Any] = {}
if raw is not None:
if isinstance(raw, Mapping):
converted = _stage_payload(
raw,
source=source,
planner=planner,
path_kind="artifacts",
)
if isinstance(converted, Mapping):
staged.update(converted)
elif isinstance(raw, (str, os.PathLike)):
converted = _stage_path_value(
raw,
key="path",
source=source,
planner=planner,
force_kind="artifacts",
)
if converted is not _DROP:
staged["path"] = converted
else:
raise ConversionError("preprocessed must be a JSON object or local path")
explicit_fields = (
"artifact_path",
"preprocessed_path",
"preprocessed_video",
"preprocessed_video_path",
"tensor_path",
)
for field in explicit_fields:
if record.get(field) is not None:
converted = _stage_path_value(
record[field],
key=field,
source=source,
planner=planner,
force_kind="artifacts",
)
if converted is not _DROP:
staged[field] = converted
if not any(_is_path_key(key) for key in staged) and source.preprocessed_roots:
root = source.preprocessed_root()
derived = _derived_files(
video_sources,
root,
_ARTIFACT_SUFFIXES,
source.media_roots.get("videos", source.media_roots.get("default")),
)
if len(derived) == 1:
staged["video_path"] = _stage_path_value(
derived[0],
key="video_path",
source=source,
planner=planner,
force_kind="artifacts",
)
elif derived:
staged["video_paths"] = _stage_path_value(
derived,
key="video_paths",
source=source,
planner=planner,
force_kind="artifacts",
)
return staged
def _choices(value: Any, source: EvaluationSource) -> list[str]:
if isinstance(value, Mapping):
value = [value[key] for key in sorted(value, key=str)]
choices = []
for item in _as_list(value):
safe = _json_safe(item, source)
if safe is _DROP:
continue
text = str(safe).strip()
if text:
choices.append(text)
return choices
def _sample_id(record: Mapping[str, Any], source: EvaluationSource, index: int) -> str:
value = _first(
record,
("sample_id", "question_id", "problem_id", "id", "uid", "index"),
)
if value is None:
return f"{source.benchmark}-{index + 1:08d}"
text = str(value).strip()
if not text:
raise ConversionError(f"source row {index + 1}: sample id must be nonempty")
if _is_private_path(text, source):
return f"{source.benchmark}-{index + 1:08d}"
return text
def _canonical_row(
record: Mapping[str, Any],
source: EvaluationSource,
planner: _AssetPlanner,
index: int,
*,
images_override: list[Any] | None = None,
videos_override: list[Any] | None = None,
) -> dict[str, Any]:
if source.adapter == "spatial_grounding":
problem_keys = (
"expression",
"normal_caption",
"caption",
"query",
"problem",
"question",
"prompt",
"input_prompt",
)
answer_keys = (
"normalized_solution",
"bbox",
_GROUND_TRUTH_BBOX_ALIAS,
"solution",
"answer",
"answers",
"ground_truth",
"oracle_label",
"oracle_labels",
_GROUND_TRUTH_ANSWER_ALIAS,
)
elif source.eval_task == "mvbench":
problem_keys = (
"problem",
"question",
"prompt",
"input_prompt",
)
answer_keys = (
"solution",
"answer",
"ground_truth",
_GROUND_TRUTH_ANSWER_ALIAS,
)
else:
problem_keys = (
"problem",
"question",
"prompt",
"input_prompt",
"expression",
"normal_caption",
"caption",
"query",
)
answer_keys = (
"answer",
"answers",
"ground_truth",
"oracle_label",
"oracle_labels",
"solution",
_GROUND_TRUTH_ANSWER_ALIAS,
"bbox",
"normalized_solution",
_GROUND_TRUTH_BBOX_ALIAS,
"span",
)
problem = _first(
record,
problem_keys,
)
answer = _first(
record,
answer_keys,
)
if source.eval_task == "mvbench" and isinstance(answer, str):
tagged_answer = re.fullmatch(
r"\s*<answer>\s*([A-Za-z])\s*</answer>\s*",
answer,
flags=re.IGNORECASE,
)
if tagged_answer is not None:
answer = tagged_answer.group(1).upper()
if problem is None:
problem = _message_content(record, "user")
if answer is None:
answer = _message_content(record, "assistant", reverse=True)
safe_problem = _json_safe(problem, source)
safe_answer = _json_safe(answer, source)
if safe_problem is _DROP:
safe_problem = None
if safe_answer is _DROP:
safe_answer = None
if (
source.eval_task == "segmentation"
and isinstance(safe_answer, str)
and not safe_answer.strip()
and isinstance(record.get("segmentation_output"), Mapping)
):
safe_answer = None
if images_override is None or videos_override is None:
staged_images, staged_videos, video_sources = _stage_media(record, source, planner)
images = staged_images if images_override is None else images_override
videos = staged_videos if videos_override is None else videos_override
else:
images = images_override
videos = videos_override
video_sources = []
subtitles, consumed_subtitle = _stage_subtitles(
record,
source,
planner,
video_sources,
)
preprocessed = _stage_preprocessed(record, source, planner, video_sources)
task_payload: dict[str, Any] = {}
parsed_answer = _structured_answer_payload(answer)
if parsed_answer is not None:
staged_answer = _stage_payload(
parsed_answer,
source=source,
planner=planner,
)
if isinstance(staged_answer, Mapping):
task_payload.update(staged_answer)
raw_payload = record.get("task_payload")
if isinstance(raw_payload, Mapping):
staged_payload = _stage_payload(raw_payload, source=source, planner=planner)
if isinstance(staged_payload, Mapping):
task_payload.update(staged_payload)
metadata: dict[str, Any] = {}
raw_metadata = record.get("metadata")
if isinstance(raw_metadata, Mapping):
staged_metadata = _stage_payload(raw_metadata, source=source, planner=planner)
if isinstance(staged_metadata, Mapping):
metadata.update(staged_metadata)
consumed = set(_BASE_CONSUMED_FIELDS)
consumed.update(
{
"artifact_path",
"preprocessed_path",
"preprocessed_video",
"preprocessed_video_path",
"tensor_path",
}
)
if consumed_subtitle:
consumed.add("subtitle")
for identifier in (
"question_id",
"problem_id",
"id",
"uid",
"index",
"data_type",
"input_prompt",
"question_type",
"task_type",
):
if record.get(identifier) is not None:
safe = _json_safe(record[identifier], source)
if safe is not _DROP:
metadata.setdefault(identifier, safe)
for raw_key, value in record.items():
if (
not isinstance(raw_key, str)
or _is_private_path(raw_key, source)
or raw_key in consumed
):
continue
if raw_key in _PAYLOAD_KEYS or _is_path_key(raw_key):
staged = (
_stage_path_value(
value,
key=raw_key,
source=source,
planner=planner,
)
if _is_path_key(raw_key)
else _stage_payload(value, source=source, planner=planner)
)
if staged is not _DROP:
task_payload.setdefault(raw_key, staged)
else:
safe = _stage_payload(value, source=source, planner=planner)
if safe is not _DROP:
metadata.setdefault(raw_key, safe)
row_evaluation = record.get("evaluation")
evaluation = _profile_without_paths(source.evaluation, source)
if not isinstance(evaluation, dict):
evaluation = {}
if isinstance(row_evaluation, Mapping):
safe_evaluation = _profile_without_paths(row_evaluation, source)
if isinstance(safe_evaluation, Mapping):
evaluation.update(safe_evaluation)
source_name = _first(record, ("source", "data_source", "dataset", "bench"))
safe_source_name = _json_safe(source_name, source)
if safe_source_name is _DROP or not str(safe_source_name or "").strip():
safe_source_name = source.benchmark
problem_type = _first(record, ("problem_type", "question_type", "task_type"))
if (
not isinstance(problem_type, str)
or not problem_type.strip()
or _is_private_path(problem_type, source)
):
problem_type = source.eval_task
row: dict[str, Any] = {
"schema_version": EVALUATION_SCHEMA_VERSION,
"eval_task": source.eval_task,
"sample_id": _sample_id(record, source, index),
"benchmark": source.benchmark,
"split": source.split,
"problem": safe_problem,
"answer": safe_answer,
"images": images,
"videos": videos,
"problem_type": problem_type,
"source": str(safe_source_name),
"family": source.family,
"task_payload": task_payload,
"evaluation": evaluation,
}
row_choices = _choices(_first(record, ("choices", "options")), source)
if row_choices:
row["choices"] = row_choices
if subtitles:
row["subtitles"] = subtitles
if preprocessed:
row["preprocessed"] = preprocessed
if metadata:
row["metadata"] = metadata
return row
def _generic_records(
source: EvaluationSource,
planner: _AssetPlanner,
) -> tuple[list[dict[str, Any]], int]:
try:
records = read_records(source.annotation_input)
except (OSError, ValueError) as error:
raise ConversionError(f"{source.annotation_input}: {error}") from error
if len(records) == 1:
container = records[0]
for field in ("annotations", "items", "questions"):
nested = container.get(field)
if isinstance(nested, list) and all(isinstance(item, Mapping) for item in nested):
records = [dict(item) for item in nested]
break
else:
row_fields = {
"problem",
"question",
"prompt",
"expression",
"normal_caption",
"caption",
}
if (
container
and not row_fields.intersection(container)
and all(isinstance(item, Mapping) for item in container.values())
):
flattened = []
for key in sorted(container, key=str):
item = dict(container[key])
item.setdefault("id", key)
flattened.append(item)
records = flattened
rows = []
for index, record in enumerate(records):
try:
rows.append(_canonical_row(record, source, planner, index))
except ConversionError as error:
raise ConversionError(f"source row {index + 1}: {error}") from error
return rows, len(records)
def _decode_base64_images(value: Any, *, context: str) -> list[bytes]:
if isinstance(value, str):
try:
parsed = ast.literal_eval(value)
except (SyntaxError, ValueError):
parsed = value
else:
parsed = value
values = _as_list(parsed)
images: list[bytes] = []
for index, item in enumerate(values):
if item is None or item == "":
continue
if isinstance(item, (bytes, bytearray, memoryview)):
images.append(bytes(item))
continue
encoded = str(item).strip()
if encoded.startswith("data:") and "," in encoded:
encoded = encoded.split(",", 1)[1]
try:
decoded = base64.b64decode(encoded, validate=True)
except (binascii.Error, ValueError) as error:
raise ConversionError(f"{context}: image {index} is not valid base64") from error
if not decoded:
raise ConversionError(f"{context}: image {index} decoded to empty bytes")
images.append(decoded)
return images
def _mmsi_records(
source: EvaluationSource,
planner: _AssetPlanner,
) -> tuple[list[dict[str, Any]], int]:
csv.field_size_limit(sys.maxsize)
try:
with source.annotation_input.open(
"r",
encoding="utf-8",
errors="replace",
newline="",
) as handle:
records = list(csv.DictReader(handle, delimiter="\t"))
except (csv.Error, OSError) as error:
raise ConversionError(f"{source.annotation_input}: invalid MMSI TSV: {error}") from error
if not records and source.expected_count:
raise ConversionError(f"{source.annotation_input}: MMSI TSV has no source rows")
rows = []
for index, record in enumerate(records):
image_bytes = _decode_base64_images(
record.get("image"),
context=f"{source.annotation_input}:{index + 2}",
)
image_paths = [
planner.add_bytes(content, kind="images", suffix=_image_suffix(content))
for content in image_bytes
]
prepared = dict(record)
prepared.pop("image", None)
rows.append(
_canonical_row(
prepared,
source,
planner,
index,
images_override=image_paths,
videos_override=[],
)
)
return rows, len(records)
def _load_parquet(source: EvaluationSource) -> list[dict[str, Any]]:
pandas_error: Exception | None = None
try:
import pandas as pd
except ImportError as error:
pandas_error = error
else:
try:
return pd.read_parquet(source.annotation_input).to_dict("records")
except ImportError as error:
pandas_error = error
try:
import pyarrow.parquet as parquet
except ImportError as error:
dependency = "pandas with a parquet engine or pyarrow"
raise ConversionError(f"parquet conversion requires {dependency}") from (
pandas_error or error
)
try:
return parquet.read_table(source.annotation_input).to_pylist()
except (OSError, ValueError) as error:
raise ConversionError(f"{source.annotation_input}: invalid parquet: {error}") from error
def _validate_official_revsi(
records: Sequence[Mapping[str, Any]],
source: EvaluationSource,
) -> None:
expected_total = sum(_REVSI_EXPECTED_GROUP_COUNTS.values())
if source.expected_count != expected_total:
return
counts = {group: 0 for group in _REVSI_EXPECTED_GROUP_COUNTS}
unknown = 0
for record in records:
question_type = str(record.get("question_type") or "").strip()
group = next(
(
prefix
for prefix in (
"object_counting",
"object_rel_direction",
"object_rel_distance",
"room_size_estimation",
)
if question_type.startswith(prefix)
),
question_type,
)
if group in counts:
counts[group] += 1
else:
unknown += 1
if counts != _REVSI_EXPECTED_GROUP_COUNTS or unknown:
observed = {**counts, **({"unknown": unknown} if unknown else {})}
raise ConversionError(
"official ReVSI all-frame split must contain "
f"{_REVSI_EXPECTED_GROUP_COUNTS}, got {observed}"
)
def _revsi_records(
source: EvaluationSource,
planner: _AssetPlanner,
) -> tuple[list[dict[str, Any]], int]:
records = _load_parquet(source)
_validate_official_revsi(records, source)
video_root = source.media_root("videos")
rows = []
for index, record in enumerate(records):
if not isinstance(record, Mapping):
raise ConversionError(
f"{source.annotation_input}: parquet row {index} is not an object"
)
scene_id = str(record.get("scene_id") or "").strip()
video_reference = _first(record, ("video_path", "video", "path"))
if video_reference is None:
if not scene_id:
raise ConversionError(
f"{source.annotation_input}: parquet row {index} has no scene_id"
)
video_reference = f"{scene_id}.mp4"
video_path, _local_path = planner.add_file(
video_reference,
kind="videos",
root=video_root,
)
rows.append(
_canonical_row(
record,
source,
planner,
index,
images_override=[],
videos_override=[video_path],
)
)
return rows, len(records)
def _mindcube_image_values(record: Mapping[str, Any]) -> list[Any]:
values = record.get("images")
if values is not None:
return _as_list(values)
fields = sorted(
(
key
for key in record
if isinstance(key, str) and re.fullmatch(r"image_?\d+", key)
),
key=lambda key: int(re.search(r"\d+", key).group()) if re.search(r"\d+", key) else 0,
)
return [record[field] for field in fields]
def _mindcube_image(
value: Any,
source: EvaluationSource,
planner: _AssetPlanner,
) -> str | None:
if value is None:
return None
if isinstance(value, Mapping):
content = value.get("bytes")
if isinstance(content, (bytes, bytearray, memoryview)):
raw = bytes(content)
return planner.add_bytes(raw, kind="images", suffix=_image_suffix(raw))
path = value.get("path")
if path:
repository_path, _local = planner.add_file(
path,
kind="images",
root=source.media_root("images"),
)
return repository_path
return None
if isinstance(value, (bytes, bytearray, memoryview)):
raw = bytes(value)
return planner.add_bytes(raw, kind="images", suffix=_image_suffix(raw))
if isinstance(value, (str, os.PathLike)):
repository_path, _local = planner.add_file(
value,
kind="images",
root=source.media_root("images"),
)
return repository_path
save = getattr(value, "save", None)
if callable(save):
buffer = io.BytesIO()
save(buffer, format="PNG")
return planner.add_bytes(buffer.getvalue(), kind="images", suffix=".png")
return None
def _mindcube_group(record: Mapping[str, Any]) -> str:
for key in ("task", "setting"):
value = str(record.get(key) or "").strip().casefold()
if value in _MINDCUBE_EXPECTED_GROUP_COUNTS:
return value
sample_id = str(
_first(record, ("id", "index", "sample_id")) or ""
).strip().casefold()
match = re.match(r"^(rotation|among|around)(?:_|$)", sample_id)
return match.group(1) if match else "unknown"
def _validate_official_mindcube(
records: Sequence[Mapping[str, Any]],
source: EvaluationSource,
) -> None:
expected_total = sum(_MINDCUBE_EXPECTED_GROUP_COUNTS.values())
if source.expected_count != expected_total:
return
counts = {group: 0 for group in _MINDCUBE_EXPECTED_GROUP_COUNTS}
unknown = 0
for record in records:
group = _mindcube_group(record)
if group in counts:
counts[group] += 1
else:
unknown += 1
if counts != _MINDCUBE_EXPECTED_GROUP_COUNTS or unknown:
observed = {**counts, **({"unknown": unknown} if unknown else {})}
raise ConversionError(
"official MindCube-Tiny must contain "
f"{_MINDCUBE_EXPECTED_GROUP_COUNTS}, got {observed}"
)
def _mindcube_records(
source: EvaluationSource,
planner: _AssetPlanner,
) -> tuple[list[dict[str, Any]], int]:
records = _load_parquet(source)
_validate_official_mindcube(records, source)
rows = []
for index, record in enumerate(records):
if not isinstance(record, Mapping):
raise ConversionError(
f"{source.annotation_input}: parquet row {index} is not an object"
)
image_paths = [
path
for value in _mindcube_image_values(record)
if (path := _mindcube_image(value, source, planner)) is not None
]
prepared = {
key: value
for key, value in record.items()
if key != "images"
and not (isinstance(key, str) and re.fullmatch(r"image_?\d+", key))
}
rows.append(
_canonical_row(
prepared,
source,
planner,
index,
images_override=image_paths,
videos_override=[],
)
)
return rows, len(records)
def _timelens_records(
source: EvaluationSource,
planner: _AssetPlanner,
) -> tuple[list[dict[str, Any]], int]:
try:
with source.annotation_input.open("r", encoding="utf-8") as handle:
payload = json.load(handle)
except (json.JSONDecodeError, OSError) as error:
raise ConversionError(
f"{source.annotation_input}: invalid TimeLens JSON: {error}"
) from error
if isinstance(payload, Mapping):
nested = [(str(video_id), info) for video_id, info in payload.items()]
elif isinstance(payload, list):
nested = []
for index, info in enumerate(payload):
if not isinstance(info, Mapping):
raise ConversionError(
f"{source.annotation_input}: TimeLens entry {index} is not an object"
)
video_id = _first(info, ("video_id", "video", "id", "name"))
nested.append((str(video_id if video_id is not None else index), info))
else:
raise ConversionError(f"{source.annotation_input}: TimeLens JSON must be an object or list")
rows: list[dict[str, Any]] = []
for video_id, raw_info in nested:
if not isinstance(raw_info, Mapping):
raise ConversionError(
f"{source.annotation_input}: TimeLens entry {video_id!r} is not an object"
)
queries = _as_list(_first(raw_info, ("queries", "sentences")))
spans = _as_list(_first(raw_info, ("spans", "timestamps")))
if (
len(queries) == 1
and len(spans) == 2
and all(isinstance(value, (int, float)) for value in spans)
):
spans = [spans]
if len(queries) != len(spans):
raise ConversionError(
f"{source.annotation_input}: TimeLens entry {video_id!r} has "
f"{len(queries)} queries but {len(spans)} spans"
)
video_reference = _first(raw_info, ("video_path", "path", "video"))
if video_reference is None or str(video_reference) == video_id:
video_reference = f"{video_id}.mp4"
common_metadata = {
key: value
for key, value in raw_info.items()
if key
not in {
"queries",
"sentences",
"spans",
"timestamps",
"video",
"video_path",
"path",
}
}
for query_index, (query, span) in enumerate(zip(queries, spans)):
query_text = re.sub(r"\s+", " ", str(query)).strip().strip(".")
prepared = {
"sample_id": f"{video_id}:{query_index:06d}",
"problem": query_text,
"answer": span,
"videos": [video_reference],
"problem_type": source.eval_task,
"source": source.benchmark,
"task_payload": {
"span": span,
**(
{"duration": raw_info["duration"]}
if raw_info.get("duration") is not None
else {}
),
},
"metadata": {
"video_id": video_id,
"query_index": query_index,
**common_metadata,
},
}
rows.append(_canonical_row(prepared, source, planner, len(rows)))
return rows, len(nested)
def convert_source(
source: EvaluationSource,
*,
hash_assets: bool = False,
) -> ConvertedSource:
"""Convert one resolved source record without writing any files."""
planner = _AssetPlanner(source, hash_assets=hash_assets)
try:
if source.adapter in GENERIC_ADAPTERS:
rows, source_rows = _generic_records(source, planner)
elif source.adapter == "mmsi":
rows, source_rows = _mmsi_records(source, planner)
elif source.adapter == "mindcube":
rows, source_rows = _mindcube_records(source, planner)
elif source.adapter == "revsi":
rows, source_rows = _revsi_records(source, planner)
elif source.adapter == "timelens":
rows, source_rows = _timelens_records(source, planner)
else:
raise ConversionError(f"unsupported adapter: {source.adapter}")
rows = sorted(rows, key=lambda row: (str(row["sample_id"]), _json_bytes(row)))
validate_evaluation_rows(
rows,
benchmark=source.benchmark,
split=source.split,
eval_task=source.eval_task,
context=f"{source.benchmark}/{source.split}",
)
except EvaluationSchemaError as error:
raise ConversionError(
f"{source.benchmark}/{source.split}: {error}"
) from error
except ConversionError as error:
raise ConversionError(
f"{source.benchmark}/{source.split}: {error}"
) from error
except (TypeError, ValueError) as error:
raise ConversionError(
f"{source.benchmark}/{source.split}: conversion failed: {error}"
) from error
return ConvertedSource(
source=source,
rows=tuple(rows),
assets=planner.assets,
source_rows=source_rows,
)
def convert_sources(
sources: Sequence[EvaluationSource],
*,
workers: int = 1,
hash_assets: bool = False,
) -> list[ConvertedSource]:
"""Convert sources deterministically, optionally in parallel by split."""
if isinstance(workers, bool) or not isinstance(workers, int) or workers <= 0:
raise ConversionError("workers must be a positive integer")
ordered = sorted(sources, key=lambda item: (item.benchmark, item.split))
if workers == 1 or len(ordered) <= 1:
return [
convert_source(source, hash_assets=hash_assets)
for source in ordered
]
with ThreadPoolExecutor(
max_workers=min(workers, len(ordered)),
thread_name_prefix="orarl-eval",
) as executor:
return list(
executor.map(
lambda source: convert_source(
source,
hash_assets=hash_assets,
),
ordered,
)
)
__all__ = [
"ConversionError",
"ConvertedSource",
"PlannedAsset",
"convert_source",
"convert_sources",
]
|