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: 30,987 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 | """Build a deterministic, leakage-filtered OraRL JSONL mixture.
Run from the OraRL directory with::
cp configs/data_sources.example.yaml ./data_sources.local.yaml
python -m orarl.data.build --config ./data_sources.local.yaml --output ./prepared/train.jsonl
"""
from __future__ import annotations
import argparse
import hashlib
import json
import os
import sys
import tempfile
from collections import Counter, defaultdict
from collections.abc import Iterable, Mapping, Sequence
from dataclasses import dataclass
from pathlib import Path
from typing import Any, Optional, Union
from .identity import media_anchors, media_leakage_tokens, prompt_identity, stable_rank
from .schema import (
SchemaError,
canonicalize_record,
is_remote_reference,
read_records,
)
PathLike = Union[str, os.PathLike[str]]
class ConfigError(ValueError):
"""Raised for an invalid source manifest."""
class ShortfallError(RuntimeError):
"""Raised when a requested train or canary count cannot be reached."""
@dataclass(frozen=True)
class SourceSpec:
name: str
input_path: Path
task: str
family: str
quota: int
media_root: Path
preserve_problem_type: bool
license: Optional[str]
url: Optional[str]
@dataclass(frozen=True)
class Candidate:
record: dict[str, Any]
spec: SourceSpec
source_index: int
identity: str
anchors: tuple[str, ...]
leakage_tokens: frozenset[str]
row_digest: str
@dataclass(frozen=True)
class Exclusions:
identities: frozenset[str]
media_tokens: frozenset[str]
rows: int
paths: tuple[Path, ...]
checksums: Mapping[str, str]
def _sha256_file(path: Path) -> str:
digest = hashlib.sha256()
with path.open("rb") as handle:
for chunk in iter(lambda: handle.read(1024 * 1024), b""):
digest.update(chunk)
return digest.hexdigest()
def _json_bytes(value: Any) -> bytes:
return json.dumps(
value,
ensure_ascii=False,
sort_keys=True,
separators=(",", ":"),
allow_nan=False,
).encode("utf-8")
def _record_digest(record: Mapping[str, Any]) -> str:
return hashlib.sha256(_json_bytes(record)).hexdigest()
def _atomic_write_jsonl(path: Path, records: Iterable[Mapping[str, Any]]) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
descriptor, temporary_name = tempfile.mkstemp(
prefix=path.name + ".",
suffix=".tmp",
dir=str(path.parent),
)
try:
with os.fdopen(descriptor, "wb") as handle:
for record in records:
handle.write(_json_bytes(record))
handle.write(b"\n")
os.replace(temporary_name, path)
except Exception:
try:
os.unlink(temporary_name)
except FileNotFoundError:
pass
raise
def _atomic_write_json(path: Path, value: Any) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
descriptor, temporary_name = tempfile.mkstemp(
prefix=path.name + ".",
suffix=".tmp",
dir=str(path.parent),
)
try:
with os.fdopen(descriptor, "w", encoding="utf-8") as handle:
json.dump(
value,
handle,
ensure_ascii=False,
sort_keys=True,
indent=2,
allow_nan=False,
)
handle.write("\n")
os.replace(temporary_name, path)
except Exception:
try:
os.unlink(temporary_name)
except FileNotFoundError:
pass
raise
def _load_yaml(path: Path) -> Any:
try:
import yaml
except ImportError as error:
raise ConfigError(
f"{path}: YAML input requires PyYAML; use a JSON manifest if it is unavailable"
) from error
try:
with path.open("r", encoding="utf-8") as handle:
return yaml.safe_load(handle)
except yaml.YAMLError as error:
raise ConfigError(f"{path}: invalid YAML: {error}") from error
def _load_config(path: Path) -> dict[str, Any]:
suffix = path.suffix.casefold()
if suffix in {".yaml", ".yml"}:
payload = _load_yaml(path)
else:
try:
with path.open("r", encoding="utf-8") as handle:
payload = json.load(handle)
except json.JSONDecodeError:
payload = _load_yaml(path)
if not isinstance(payload, dict):
raise ConfigError(f"{path}: source manifest must be a mapping")
return payload
def _local_path(base: Path, value: Any, label: str) -> Path:
if not isinstance(value, str) or not value.strip():
raise ConfigError(f"{label} must be a nonempty local path")
if is_remote_reference(value):
raise ConfigError(f"{label} must be local; URL fields are metadata only")
path = Path(os.path.expanduser(value.strip()))
if not path.is_absolute():
path = base / path
return Path(os.path.abspath(os.path.normpath(str(path))))
def _optional_text(value: Any, label: str) -> Optional[str]:
if value is None:
return None
if not isinstance(value, str) or not value.strip():
raise ConfigError(f"{label} must be a nonempty string when supplied")
return value.strip()
def _integer(value: Any, label: str, minimum: int = 0) -> int:
if isinstance(value, bool):
raise ConfigError(f"{label} must be an integer")
try:
parsed = int(value)
except (TypeError, ValueError) as error:
raise ConfigError(f"{label} must be an integer") from error
if parsed != value and not (isinstance(value, str) and str(parsed) == value.strip()):
raise ConfigError(f"{label} must be an integer")
if parsed < minimum:
raise ConfigError(f"{label} must be at least {minimum}")
return parsed
def _boolean(value: Any, label: str, default: bool) -> bool:
if value is None:
return default
if not isinstance(value, bool):
raise ConfigError(f"{label} must be true or false")
return value
def _source_specs(config_path: Path, config: Mapping[str, Any]) -> list[SourceSpec]:
raw_sources = config.get("sources")
if not isinstance(raw_sources, list) or not raw_sources:
raise ConfigError(f"{config_path}: sources must be a nonempty list")
specs: list[SourceSpec] = []
names: set[str] = set()
for index, raw in enumerate(raw_sources):
label = f"{config_path}: sources[{index}]"
if not isinstance(raw, Mapping):
raise ConfigError(f"{label} must be a mapping")
raw_input = raw.get("input", raw.get("path"))
input_path = _local_path(config_path.parent, raw_input, f"{label}.input")
raw_name = raw.get("name", raw.get("source"))
name = _optional_text(raw_name, f"{label}.name")
if name is None:
name = input_path.stem
if name in names:
raise ConfigError(f"{label}.name is duplicated: {name}")
names.add(name)
task = _optional_text(raw.get("task"), f"{label}.task")
family = _optional_text(raw.get("family"), f"{label}.family")
if task is None or family is None:
raise ConfigError(f"{label} must define task and family")
quota = _integer(raw.get("quota"), f"{label}.quota")
media_root = _local_path(
config_path.parent,
raw.get("media_root"),
f"{label}.media_root",
)
if not input_path.is_file():
raise ConfigError(f"{label}.input does not exist: {input_path}")
specs.append(
SourceSpec(
name=name,
input_path=input_path,
task=task,
family=family,
quota=quota,
media_root=media_root,
preserve_problem_type=_boolean(
raw.get("preserve_problem_type"),
f"{label}.preserve_problem_type",
False,
),
license=_optional_text(raw.get("license"), f"{label}.license"),
url=_optional_text(raw.get("url"), f"{label}.url"),
)
)
return specs
def _benchmark_paths(
config_path: Path,
config: Mapping[str, Any],
command_line_paths: Sequence[PathLike],
) -> list[Path]:
raw_paths = config.get(
"benchmark_excludes",
config.get("benchmark_exclude", []),
)
if raw_paths is None:
raw_paths = []
if isinstance(raw_paths, (str, Mapping)):
raw_paths = [raw_paths]
if not isinstance(raw_paths, list):
raise ConfigError(f"{config_path}: benchmark_excludes must be a list")
paths: list[Path] = []
for index, raw in enumerate(raw_paths):
value = raw.get("path") if isinstance(raw, Mapping) else raw
paths.append(
_local_path(
config_path.parent,
value,
f"{config_path}: benchmark_excludes[{index}]",
)
)
for index, raw in enumerate(command_line_paths):
paths.append(_local_path(Path.cwd(), str(raw), f"--benchmark-exclude[{index}]"))
unique: list[Path] = []
seen: set[str] = set()
for path in paths:
key = os.path.normcase(str(path))
if key not in seen:
seen.add(key)
if not path.is_file():
raise ConfigError(f"benchmark exclusion file does not exist: {path}")
unique.append(path)
return unique
def _loose_benchmark_record(raw: Mapping[str, Any], root: Path) -> tuple[dict[str, Any], bool]:
has_problem = any(
isinstance(raw.get(key), str) and bool(str(raw[key]).strip())
for key in ("problem", "question", "prompt")
)
working = dict(raw)
if not has_problem:
working["problem"] = "media exclusion"
working["answer"] = "excluded"
if not working.get("problem_type"):
working["problem_type"] = "benchmark"
if not any(
working.get(key)
for key in ("images", "videos", "image", "video", "image_path", "video_path")
):
generic_media = raw.get("media", raw.get("media_path", raw.get("path")))
if generic_media:
working["videos"] = generic_media
return (
canonicalize_record(
working,
source="benchmark",
media_root=root,
require_media=False,
context="benchmark exclusion",
),
has_problem and bool(raw.get("problem_type")),
)
def _load_exclusions(paths: Sequence[Path]) -> Exclusions:
identities: set[str] = set()
tokens: set[str] = set()
rows = 0
checksums: dict[str, str] = {}
for path in paths:
checksums[str(path)] = _sha256_file(path)
for raw in read_records(path):
rows += 1
explicit_identity = raw.get("prompt_identity", raw.get("identity"))
if isinstance(explicit_identity, str) and explicit_identity.strip():
identities.add(explicit_identity.strip())
try:
record, has_full_identity = _loose_benchmark_record(raw, path.parent)
except SchemaError:
if explicit_identity:
continue
raise
tokens.update(media_leakage_tokens(record))
if has_full_identity:
identities.add(prompt_identity(record))
return Exclusions(
identities=frozenset(identities),
media_tokens=frozenset(tokens),
rows=rows,
paths=tuple(paths),
checksums=checksums,
)
def _load_candidates(
specs: Sequence[SourceSpec],
require_media: bool,
) -> tuple[list[Candidate], dict[str, int], dict[str, str]]:
candidates: list[Candidate] = []
input_counts: dict[str, int] = {}
checksums: dict[str, str] = {}
for spec in specs:
raw_records = read_records(spec.input_path)
input_counts[spec.name] = len(raw_records)
checksums[spec.name] = _sha256_file(spec.input_path)
for index, raw in enumerate(raw_records):
prepared = dict(raw)
if (
not spec.preserve_problem_type
or not str(prepared.get("problem_type") or "").strip()
):
prepared["problem_type"] = spec.task
context = f"{spec.input_path}: record {index + 1}"
record = canonicalize_record(
prepared,
source=spec.name,
family=spec.family,
media_root=spec.media_root,
require_media=require_media,
context=context,
)
try:
digest = _record_digest(record)
except (TypeError, ValueError) as error:
raise SchemaError(f"{context}: record is not strict JSON: {error}") from error
candidates.append(
Candidate(
record=record,
spec=spec,
source_index=index,
identity=prompt_identity(record),
anchors=media_anchors(record),
leakage_tokens=media_leakage_tokens(record),
row_digest=digest,
)
)
return candidates, input_counts, checksums
def _candidate_order(candidate: Candidate, seed: int, namespace: str) -> tuple[int, str, int]:
key = f"{candidate.identity}|{candidate.row_digest}"
return (
stable_rank(seed, namespace, key),
candidate.row_digest,
candidate.source_index,
)
def _select_train(
specs: Sequence[SourceSpec],
candidates: Sequence[Candidate],
seed: int,
max_prompts_per_media: int,
) -> tuple[list[Candidate], dict[str, int], dict[str, int]]:
grouped: dict[str, list[Candidate]] = defaultdict(list)
for candidate in candidates:
grouped[candidate.spec.name].append(candidate)
unique_capacity = {
spec.name: len({candidate.identity for candidate in grouped.get(spec.name, [])})
for spec in specs
}
spec_order = sorted(
specs,
key=lambda spec: (
unique_capacity[spec.name] - spec.quota,
spec.name.casefold(),
),
)
selected: list[Candidate] = []
selected_identities: set[str] = set()
anchor_counts: Counter[str] = Counter()
cap_rejections: Counter[str] = Counter()
selected_counts: Counter[str] = Counter()
for spec in spec_order:
ordered = sorted(
grouped.get(spec.name, []),
key=lambda candidate: _candidate_order(
candidate,
seed,
f"train:{spec.name}",
),
)
for candidate in ordered:
if selected_counts[spec.name] >= spec.quota:
break
if candidate.identity in selected_identities:
continue
if any(anchor_counts[anchor] >= max_prompts_per_media for anchor in candidate.anchors):
cap_rejections[spec.name] += 1
continue
selected.append(candidate)
selected_counts[spec.name] += 1
selected_identities.add(candidate.identity)
anchor_counts.update(candidate.anchors)
shortfalls = {
spec.name: spec.quota - selected_counts[spec.name]
for spec in specs
if selected_counts[spec.name] < spec.quota
}
return selected, shortfalls, dict(cap_rejections)
def _select_canary(
candidates: Sequence[Candidate],
train: Sequence[Candidate],
canary_size: int,
seed: int,
max_prompts_per_media: int,
) -> tuple[list[Candidate], dict[str, int]]:
train_identities = {candidate.identity for candidate in train}
train_media_tokens: set[str] = set()
for candidate in train:
train_media_tokens.update(candidate.leakage_tokens)
ordered = sorted(
candidates,
key=lambda candidate: _candidate_order(candidate, seed, "canary"),
)
canary: list[Candidate] = []
canary_identities: set[str] = set()
canary_anchor_counts: Counter[str] = Counter()
stats: Counter[str] = Counter()
for candidate in ordered:
if len(canary) >= canary_size:
break
if candidate.identity in train_identities:
stats["train_identity"] += 1
continue
if candidate.identity in canary_identities:
stats["duplicate_identity"] += 1
continue
if candidate.leakage_tokens & train_media_tokens:
stats["train_media"] += 1
continue
if any(
canary_anchor_counts[anchor] >= max_prompts_per_media for anchor in candidate.anchors
):
stats["media_cap"] += 1
continue
canary.append(candidate)
canary_identities.add(candidate.identity)
canary_anchor_counts.update(candidate.anchors)
return canary, dict(stats)
def _sorted_output(candidates: Sequence[Candidate], seed: int, namespace: str) -> list[Candidate]:
return sorted(
candidates,
key=lambda candidate: _candidate_order(candidate, seed, namespace),
)
def _counts(candidates: Sequence[Candidate]) -> dict[str, Any]:
by_task = Counter(str(candidate.record["problem_type"]) for candidate in candidates)
by_source = Counter(str(candidate.record["source"]) for candidate in candidates)
by_family = Counter(str(candidate.record.get("family") or "") for candidate in candidates)
return {
"rows": len(candidates),
"by_task": dict(sorted(by_task.items())),
"by_source": dict(sorted(by_source.items())),
"by_family": dict(sorted(by_family.items())),
}
def _default_canary_path(output: Path) -> Path:
suffix = output.suffix or ".jsonl"
stem = output.name[: -len(output.suffix)] if output.suffix else output.name
return output.with_name(stem + ".canary" + suffix)
def _default_manifest_path(output: Path) -> Path:
stem = output.name[: -len(output.suffix)] if output.suffix else output.name
return output.with_name(stem + ".manifest.json")
def _output_path(value: PathLike) -> Path:
path = Path(os.path.expanduser(str(value)))
return Path(os.path.abspath(os.path.normpath(str(path))))
def _check_output_paths(
paths: Sequence[Path],
overwrite: bool,
protected_paths: Sequence[Path],
) -> None:
normalized = [os.path.normcase(str(path)) for path in paths]
if len(normalized) != len(set(normalized)):
raise ConfigError("train, canary, and manifest outputs must be different paths")
protected = {os.path.normcase(str(path)) for path in protected_paths}
collisions = [
str(path)
for path, normalized_path in zip(paths, normalized)
if normalized_path in protected
]
if collisions:
raise ConfigError(
"output paths cannot replace configs, sources, or exclusions: " + ", ".join(collisions)
)
if not overwrite:
existing = [str(path) for path in paths if path.exists()]
if existing:
raise FileExistsError(
"refusing to overwrite existing output(s): "
+ ", ".join(existing)
+ "; pass --overwrite"
)
def _output_leakage(
train: Sequence[Candidate],
canary: Sequence[Candidate],
exclusions: Exclusions,
) -> dict[str, int]:
train_ids = {candidate.identity for candidate in train}
canary_ids = {candidate.identity for candidate in canary}
train_tokens: set[str] = set()
canary_tokens: set[str] = set()
for candidate in train:
train_tokens.update(candidate.leakage_tokens)
for candidate in canary:
canary_tokens.update(candidate.leakage_tokens)
return {
"benchmark_identity": len((train_ids | canary_ids) & exclusions.identities),
"benchmark_media": len((train_tokens | canary_tokens) & exclusions.media_tokens),
"train_canary_identity": len(train_ids & canary_ids),
"train_canary_media": len(train_tokens & canary_tokens),
}
def build_dataset(
config_path: PathLike,
output_path: PathLike,
*,
canary_output: Optional[PathLike] = None,
manifest_output: Optional[PathLike] = None,
seed: Optional[int] = None,
target: Optional[int] = None,
canary_size: Optional[int] = None,
max_prompts_per_media: Optional[int] = None,
require_media: Optional[bool] = None,
benchmark_excludes: Sequence[PathLike] = (),
allow_shortfall: Optional[bool] = None,
overwrite: bool = False,
) -> dict[str, Any]:
"""Build train, canary, and audit manifest files."""
config_file = _output_path(config_path)
if not config_file.is_file():
raise ConfigError(f"source manifest does not exist: {config_file}")
config = _load_config(config_file)
specs = _source_specs(config_file, config)
effective_seed = _integer(
seed if seed is not None else config.get("seed", 42),
"seed",
)
effective_canary_size = _integer(
canary_size if canary_size is not None else config.get("canary_size", 0),
"canary_size",
)
effective_cap = _integer(
(
max_prompts_per_media
if max_prompts_per_media is not None
else config.get(
"max_prompts_per_media",
config.get("max_prompts_per_anchor", 1),
)
),
"max_prompts_per_media",
minimum=1,
)
effective_require_media = (
require_media
if require_media is not None
else _boolean(config.get("require_media"), "require_media", False)
)
effective_allow_shortfall = (
allow_shortfall
if allow_shortfall is not None
else _boolean(config.get("allow_shortfall"), "allow_shortfall", False)
)
quota_total = sum(spec.quota for spec in specs)
effective_target = _integer(
target if target is not None else config.get("target", quota_total),
"target",
)
if effective_target != quota_total:
raise ConfigError(
f"target ({effective_target}) must equal the sum of source quotas ({quota_total})"
)
output = _output_path(output_path)
canary_path = (
_output_path(canary_output) if canary_output is not None else _default_canary_path(output)
)
manifest_path = (
_output_path(manifest_output)
if manifest_output is not None
else _default_manifest_path(output)
)
benchmark_paths = _benchmark_paths(config_file, config, benchmark_excludes)
protected_paths = [
config_file,
*(spec.input_path for spec in specs),
*benchmark_paths,
]
_check_output_paths(
(output, canary_path, manifest_path),
overwrite,
protected_paths,
)
exclusions = _load_exclusions(benchmark_paths)
all_candidates, input_counts, source_checksums = _load_candidates(
specs,
require_media=effective_require_media,
)
identity_counts = Counter(candidate.identity for candidate in all_candidates)
duplicate_rows = sum(count - 1 for count in identity_counts.values())
duplicate_groups = sum(count > 1 for count in identity_counts.values())
eligible: list[Candidate] = []
identity_leak_rows = 0
media_leak_rows = 0
excluded_rows = 0
for candidate in all_candidates:
identity_hit = candidate.identity in exclusions.identities
media_hit = bool(candidate.leakage_tokens & exclusions.media_tokens)
identity_leak_rows += int(identity_hit)
media_leak_rows += int(media_hit)
if identity_hit or media_hit:
excluded_rows += 1
else:
eligible.append(candidate)
train, source_shortfalls, cap_rejections = _select_train(
specs,
eligible,
effective_seed,
effective_cap,
)
if source_shortfalls and not effective_allow_shortfall:
detail = ", ".join(
f"{name} short by {count}" for name, count in sorted(source_shortfalls.items())
)
raise ShortfallError(f"source quota shortfall: {detail}")
canary, canary_rejections = _select_canary(
eligible,
train,
effective_canary_size,
effective_seed,
effective_cap,
)
canary_shortfall = effective_canary_size - len(canary)
if canary_shortfall and not effective_allow_shortfall:
raise ShortfallError(
f"canary shortfall: requested {effective_canary_size}, got {len(canary)}"
)
train = _sorted_output(train, effective_seed, "train-output")
canary = _sorted_output(canary, effective_seed, "canary-output")
output_leakage = _output_leakage(train, canary, exclusions)
if any(output_leakage.values()):
raise RuntimeError(f"internal leakage invariant failed: {output_leakage}")
_atomic_write_jsonl(output, (candidate.record for candidate in train))
_atomic_write_jsonl(canary_path, (candidate.record for candidate in canary))
source_details = []
selected_by_source = Counter(candidate.spec.name for candidate in train)
for spec in sorted(specs, key=lambda value: value.name.casefold()):
source_details.append(
{
"name": spec.name,
"input": str(spec.input_path),
"media_root": str(spec.media_root),
"task": spec.task,
"family": spec.family,
"quota": spec.quota,
"selected": selected_by_source[spec.name],
"shortfall": spec.quota - selected_by_source[spec.name],
"preserve_problem_type": spec.preserve_problem_type,
"license": spec.license,
"url": spec.url,
"input_sha256": source_checksums[spec.name],
}
)
manifest: dict[str, Any] = {
"schema_version": 1,
"seed": effective_seed,
"target": effective_target,
"actual_train_rows": len(train),
"canary_target": effective_canary_size,
"actual_canary_rows": len(canary),
"allow_shortfall": effective_allow_shortfall,
"media": {
"require_existing": effective_require_media,
"max_prompts_per_anchor": effective_cap,
},
"files": {
"train": str(output),
"canary": str(canary_path),
"manifest": str(manifest_path),
},
"checksums": {
"algorithm": "sha256",
"config": _sha256_file(config_file),
"train": _sha256_file(output),
"canary": _sha256_file(canary_path),
"sources": dict(sorted(source_checksums.items())),
"benchmark_excludes": dict(sorted(exclusions.checksums.items())),
},
"counts": {
"input_rows": len(all_candidates),
"input_by_source": dict(sorted(input_counts.items())),
"eligible_rows": len(eligible),
"train": _counts(train),
"canary": _counts(canary),
},
"duplicates": {
"candidate_prompt_identity_groups": duplicate_groups,
"candidate_prompt_rows": duplicate_rows,
"train_prompt_rows": len(train) - len({candidate.identity for candidate in train}),
"canary_prompt_rows": len(canary) - len({candidate.identity for candidate in canary}),
},
"leakage": {
"benchmark_reference_rows": exclusions.rows,
"candidate_identity_rows_excluded": identity_leak_rows,
"candidate_media_rows_excluded": media_leak_rows,
"candidate_rows_excluded": excluded_rows,
"output": output_leakage,
},
"selection": {
"source_shortfalls": dict(sorted(source_shortfalls.items())),
"canary_shortfall": canary_shortfall,
"media_cap_rejections_by_source": dict(sorted(cap_rejections.items())),
"canary_rejections": dict(sorted(canary_rejections.items())),
},
"sources": source_details,
"benchmark_excludes": [str(path) for path in exclusions.paths],
}
_atomic_write_json(manifest_path, manifest)
return manifest
def build_parser() -> argparse.ArgumentParser:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--config", required=True, help="YAML or JSON source manifest")
parser.add_argument("--output", required=True, help="Training JSONL output")
parser.add_argument("--canary-output", default=None)
parser.add_argument("--manifest-output", default=None)
parser.add_argument("--seed", type=int, default=None)
parser.add_argument("--target", type=int, default=None)
parser.add_argument("--canary-size", type=int, default=None)
parser.add_argument(
"--max-prompts-per-media",
"--max-prompts-per-anchor",
dest="max_prompts_per_media",
type=int,
default=None,
)
parser.add_argument(
"--require-media",
action=argparse.BooleanOptionalAction,
default=None,
help="Require every normalized media path to exist",
)
parser.add_argument(
"--require-existing-media",
dest="require_media",
action="store_true",
help=argparse.SUPPRESS,
)
parser.add_argument(
"--benchmark-exclude",
action="append",
default=[],
help="JSON or JSONL benchmark identities/media to exclude; repeatable",
)
parser.add_argument(
"--allow-shortfall",
action="store_true",
default=None,
help="Write available rows instead of failing a source or canary shortfall",
)
parser.add_argument("--overwrite", action="store_true")
return parser
def main(argv: Optional[Sequence[str]] = None) -> int:
args = build_parser().parse_args(argv)
try:
manifest = build_dataset(
args.config,
args.output,
canary_output=args.canary_output,
manifest_output=args.manifest_output,
seed=args.seed,
target=args.target,
canary_size=args.canary_size,
max_prompts_per_media=args.max_prompts_per_media,
require_media=args.require_media,
benchmark_excludes=args.benchmark_exclude,
allow_shortfall=args.allow_shortfall,
overwrite=args.overwrite,
)
except (ConfigError, FileExistsError, OSError, SchemaError, ShortfallError) as error:
print(f"[FATAL] {error}", file=sys.stderr)
return 2
print(
"OraRL data prepared: "
f"train={manifest['actual_train_rows']} "
f"canary={manifest['actual_canary_rows']} "
f"manifest={manifest['files']['manifest']}"
)
return 0
if __name__ == "__main__":
raise SystemExit(main())
|