File size: 35,333 Bytes
d82db76 | 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 | """Validation gates for defensible OCR quality claims."""
from __future__ import annotations
import json
import math
import random
from dataclasses import dataclass, field
from pathlib import Path
from typing import Any
from .errors import ValidationError
from .manifest import ManifestEntry, load_manifest, sha256_file, sha256_text
LOWER_IS_BETTER = {
"cer",
"wer",
"text_cer",
"text_wer",
"document_reconstruction_cer",
"document_reconstruction_wer",
"figure_count_error",
"figure_caption_cer",
}
HIGHER_IS_BETTER = {
"quality_score",
"document_reconstruction_similarity",
"figure_detection_precision",
"figure_detection_recall",
"figure_detection_f1",
"reading_order_pair_accuracy",
"table_cell_precision",
"table_cell_recall",
"table_cell_f1",
"table_exact_match",
"table_structure_similarity",
"table_teds_structure",
"table_teds_text",
}
@dataclass(slots=True)
class ClaimGate:
baseline: str
metric: str
direction: str
slice: str | None = None
slices: list[str] = field(default_factory=list)
min_absolute_margin: float = 0.0
min_relative_improvement: float = 0.0
min_pairs: int = 1
min_pair_win_rate: float | None = None
max_p_value: float | None = None
require_ci_above_zero: bool = True
@classmethod
def from_dict(cls, data: dict[str, Any]) -> "ClaimGate":
metric = str(data.get("metric") or "")
direction = str(data.get("direction") or metric_direction(metric))
if direction not in {"lower", "higher"}:
raise ValidationError(f"gate direction must be lower or higher: {direction}")
baseline = str(data.get("baseline") or "")
if not baseline:
raise ValidationError("gate baseline is required")
if not metric:
raise ValidationError("gate metric is required")
slices_payload = data.get("slices")
if slices_payload is None:
slices = [str(data["slice"])] if data.get("slice") else []
elif isinstance(slices_payload, list):
slices = [str(item) for item in slices_payload if str(item)]
else:
raise ValidationError("gate slices must be a list when provided")
min_pair_win_rate = _optional_rate(data.get("min_pair_win_rate"), "min_pair_win_rate")
return cls(
baseline=baseline,
metric=metric,
direction=direction,
slice=data.get("slice"),
slices=slices,
min_absolute_margin=float(data.get("min_absolute_margin", 0.0)),
min_relative_improvement=float(data.get("min_relative_improvement", 0.0)),
min_pairs=int(data.get("min_pairs", 1)),
min_pair_win_rate=min_pair_win_rate,
max_p_value=float(data["max_p_value"]) if data.get("max_p_value") is not None else None,
require_ci_above_zero=bool(data.get("require_ci_above_zero", True)),
)
def to_dict(self) -> dict[str, Any]:
return {
"baseline": self.baseline,
"metric": self.metric,
"direction": self.direction,
"slice": self.slice,
"slices": self.slices,
"min_absolute_margin": self.min_absolute_margin,
"min_relative_improvement": self.min_relative_improvement,
"min_pairs": self.min_pairs,
"min_pair_win_rate": self.min_pair_win_rate,
"max_p_value": self.max_p_value,
"require_ci_above_zero": self.require_ci_above_zero,
}
@dataclass(slots=True)
class ValidationConfig:
candidate: str = "ours"
min_samples: int = 30
required_slices: dict[str, int] = field(default_factory=dict)
gates: list[ClaimGate] = field(default_factory=list)
require_full_eval_manifest_coverage: bool = False
bootstrap_iterations: int = 1000
bootstrap_confidence: float = 0.95
random_seed: int = 13
@classmethod
def from_path(cls, path: str | Path | None) -> "ValidationConfig":
if path is None:
return default_validation_config()
try:
payload = json.loads(Path(path).read_text(encoding="utf-8"))
except json.JSONDecodeError as exc:
raise ValidationError(f"Invalid validation config JSON {path}: {exc}") from exc
if not isinstance(payload, dict):
raise ValidationError("validation config must be a JSON object")
return cls(
candidate=str(payload.get("candidate", "ours")),
min_samples=int(payload.get("min_samples", 30)),
required_slices={str(key): int(value) for key, value in dict(payload.get("required_slices", {})).items()},
gates=[ClaimGate.from_dict(item) for item in payload.get("gates", [])],
require_full_eval_manifest_coverage=bool(payload.get("require_full_eval_manifest_coverage", False)),
bootstrap_iterations=int(payload.get("bootstrap_iterations", 1000)),
bootstrap_confidence=float(payload.get("bootstrap_confidence", 0.95)),
random_seed=int(payload.get("random_seed", 13)),
)
@dataclass(slots=True)
class LeakageReport:
train_manifests: list[str]
eval_manifest: str | None
overlaps: dict[str, list[str]]
warnings: dict[str, list[str]] = field(default_factory=dict)
@property
def passed(self) -> bool:
return not any(self.overlaps.values())
def to_dict(self) -> dict[str, Any]:
return {
"passed": self.passed,
"train_manifests": self.train_manifests,
"eval_manifest": self.eval_manifest,
"overlaps": self.overlaps,
"warnings": self.warnings,
}
@dataclass(slots=True)
class GateResult:
gate: ClaimGate
passed: bool
pairs: int
candidate_mean: float | None
baseline_mean: float | None
absolute_improvement: float | None
relative_improvement: float | None
improvement_ci_low: float | None
improvement_ci_high: float | None
win_count: int
loss_count: int
tie_count: int
win_rate: float | None
pair_win_rate: float | None
sign_test_p_value: float | None
reason: str | None = None
def to_dict(self) -> dict[str, Any]:
return {
"gate": self.gate.to_dict(),
"passed": self.passed,
"pairs": self.pairs,
"candidate_mean": self.candidate_mean,
"baseline_mean": self.baseline_mean,
"absolute_improvement": self.absolute_improvement,
"relative_improvement": self.relative_improvement,
"improvement_ci_low": self.improvement_ci_low,
"improvement_ci_high": self.improvement_ci_high,
"win_count": self.win_count,
"loss_count": self.loss_count,
"tie_count": self.tie_count,
"win_rate": self.win_rate,
"pair_win_rate": self.pair_win_rate,
"sign_test_p_value": self.sign_test_p_value,
"reason": self.reason,
}
@dataclass(slots=True)
class ValidationReport:
claim_status: str
passed: bool
sample_count: int
slice_counts: dict[str, int]
aggregate_metrics: dict[str, dict[str, dict[str, float | int]]]
missing_requirements: list[str]
leakage: LeakageReport
gates: list[GateResult]
provenance: dict[str, Any] = field(default_factory=dict)
def to_dict(self) -> dict[str, Any]:
return {
"claim_status": self.claim_status,
"passed": self.passed,
"sample_count": self.sample_count,
"slice_counts": self.slice_counts,
"aggregate_metrics": self.aggregate_metrics,
"missing_requirements": self.missing_requirements,
"leakage": self.leakage.to_dict(),
"gates": [gate.to_dict() for gate in self.gates],
"provenance": self.provenance,
}
def validate_claim(
benchmark_report: str | Path,
output_dir: str | Path,
*,
config_path: str | Path | None = None,
eval_manifest: str | Path | None = None,
train_manifests: list[str | Path] | None = None,
candidate_model_config: str | Path | None = None,
fallback_model_config: str | Path | None = None,
) -> ValidationReport:
config = ValidationConfig.from_path(config_path)
results = _load_benchmark_results(benchmark_report)
eval_entries = load_manifest(eval_manifest) if eval_manifest else []
enriched = [_enrich_result(item, eval_entries) for item in results]
candidate_rows = [item for item in enriched if item["baseline"] == config.candidate and item["status"] == "ok"]
sample_count = len({sample_key(item) for item in candidate_rows})
slice_counts = _slice_counts(candidate_rows)
leakage = check_leakage(train_manifests or [], eval_manifest)
missing: list[str] = []
if sample_count < config.min_samples:
missing.append(f"candidate sample count {sample_count} < required {config.min_samples}")
for slice_name, required_count in config.required_slices.items():
actual = slice_counts.get(slice_name, 0)
if actual < required_count:
missing.append(f"slice {slice_name!r} count {actual} < required {required_count}")
missing.extend(_full_eval_manifest_coverage_failures(enriched, config, eval_entries))
if not leakage.passed:
missing.append("train/eval leakage detected")
gate_results = [_evaluate_gate(enriched, config, gate) for gate in config.gates]
for gate_result in gate_results:
if not gate_result.passed:
missing.append(gate_result.reason or f"gate failed: {gate_result.gate.metric} vs {gate_result.gate.baseline}")
passed = not missing and bool(config.gates)
report = ValidationReport(
claim_status="validated" if passed else "not_validated",
passed=passed,
sample_count=sample_count,
slice_counts=slice_counts,
aggregate_metrics=_aggregate_metrics(enriched),
missing_requirements=missing,
leakage=leakage,
gates=gate_results,
provenance=_validation_provenance(
benchmark_report,
config_path,
eval_manifest,
train_manifests or [],
candidate_model_config,
fallback_model_config,
),
)
_write_validation_report(report, Path(output_dir))
return report
def check_leakage(train_manifests: list[str | Path], eval_manifest: str | Path | None) -> LeakageReport:
if eval_manifest is None or not train_manifests:
return LeakageReport([str(item) for item in train_manifests], str(eval_manifest) if eval_manifest else None, {})
eval_entries = load_manifest(eval_manifest)
train_entries: list[ManifestEntry] = []
for manifest in train_manifests:
train_entries.extend(load_manifest(manifest))
overlaps = {
"sample_id": sorted({entry.sample_id for entry in train_entries} & {entry.sample_id for entry in eval_entries}),
"image_sha256": _hash_overlap(train_entries, eval_entries, "sha256"),
"sample_sha256": _sample_hash_overlap(train_entries, eval_entries),
"text_sha256": _text_hash_overlap(train_entries, eval_entries),
"image_path": sorted({entry.image_path for entry in train_entries} & {entry.image_path for entry in eval_entries}),
}
warnings = {"text_sha256": overlaps.pop("text_sha256")}
return LeakageReport([str(item) for item in train_manifests], str(eval_manifest), overlaps, warnings)
def _validation_provenance(
benchmark_report: str | Path,
config_path: str | Path | None,
eval_manifest: str | Path | None,
train_manifests: list[str | Path],
candidate_model_config: str | Path | None,
fallback_model_config: str | Path | None,
) -> dict[str, Any]:
benchmark_path = Path(benchmark_report)
payload: dict[str, Any] = {
"benchmark_report": {
"path": str(benchmark_path),
"sha256": _sha256_if_file(benchmark_path),
},
"validation_config": None,
"eval_manifest": None,
"candidate_model_config": None,
"fallback_model_config": None,
"train_manifests": [],
}
if config_path is not None:
config = Path(config_path)
payload["validation_config"] = {"path": str(config), "sha256": _sha256_if_file(config)}
if eval_manifest is not None:
eval_path = Path(eval_manifest)
payload["eval_manifest"] = {"path": str(eval_path), "sha256": _sha256_if_file(eval_path)}
if candidate_model_config is not None:
candidate_path = Path(candidate_model_config)
payload["candidate_model_config"] = {"path": str(candidate_path), "sha256": _sha256_if_file(candidate_path)}
if fallback_model_config is not None:
fallback_path = Path(fallback_model_config)
payload["fallback_model_config"] = {"path": str(fallback_path), "sha256": _sha256_if_file(fallback_path)}
for manifest in train_manifests:
path = Path(manifest)
payload["train_manifests"].append({"path": str(path), "sha256": _sha256_if_file(path)})
return payload
def _sha256_if_file(path: Path) -> str | None:
if not path.exists() or not path.is_file():
return None
return sha256_file(path)
def _optional_rate(value: object, field_name: str) -> float | None:
if value is None:
return None
rate = float(value)
if rate < 0 or rate > 1:
raise ValidationError(f"{field_name} must be between 0 and 1")
return rate
def default_validation_config() -> ValidationConfig:
return ValidationConfig(
candidate="ours",
min_samples=30,
required_slices={"nepali": 10, "english": 10, "table": 5, "reading_order": 5, "figure": 5},
require_full_eval_manifest_coverage=False,
gates=[
ClaimGate(baseline="tesseract", metric="cer", direction="lower", slice="nepali", min_relative_improvement=0.05, min_pair_win_rate=0.55),
ClaimGate(baseline="tesseract", metric="wer", direction="lower", slice="nepali", min_relative_improvement=0.02, min_pair_win_rate=0.55),
ClaimGate(baseline="stock-paddle", metric="cer", direction="lower", slice="nepali", min_relative_improvement=0.02, min_pair_win_rate=0.55),
ClaimGate(baseline="stock-paddle", metric="cer", direction="lower", slice="english", min_relative_improvement=0.0, require_ci_above_zero=False),
ClaimGate(baseline="stock-paddle", metric="wer", direction="lower", slice="english", min_relative_improvement=0.0, require_ci_above_zero=False),
ClaimGate(baseline="glm-ocr", metric="cer", direction="lower", slice="nepali", min_relative_improvement=0.02, min_pair_win_rate=0.55),
ClaimGate(baseline="surya", metric="table_cell_f1", direction="higher", slice="table", min_absolute_margin=0.01, min_pair_win_rate=0.55),
ClaimGate(baseline="surya", metric="table_teds_text", direction="higher", slice="table", min_absolute_margin=0.01, min_pair_win_rate=0.55),
ClaimGate(baseline="paddleocr-vl", metric="reading_order_pair_accuracy", direction="higher", slice="reading_order", min_absolute_margin=0.01, min_pair_win_rate=0.55),
ClaimGate(baseline="paddleocr-vl", metric="document_reconstruction_similarity", direction="higher", slice="figure", min_absolute_margin=0.01, min_pair_win_rate=0.55),
ClaimGate(baseline="surya", metric="figure_detection_f1", direction="higher", slice="figure", min_absolute_margin=0.0, require_ci_above_zero=False),
ClaimGate(baseline="surya", metric="figure_caption_cer", direction="lower", slice="figure", min_absolute_margin=0.01, min_pair_win_rate=0.55),
],
)
def metric_direction(metric: str) -> str:
if metric in LOWER_IS_BETTER:
return "lower"
if metric in HIGHER_IS_BETTER:
return "higher"
raise ValidationError(f"Unknown metric direction for {metric!r}; specify direction explicitly")
def sample_key(result: dict[str, Any]) -> str:
return str(result.get("sample_id") or result.get("input_path") or "")
def _load_benchmark_results(path: str | Path) -> list[dict[str, Any]]:
payload = json.loads(Path(path).read_text(encoding="utf-8"))
if not isinstance(payload, list):
raise ValidationError("benchmark report must be a JSON list")
rows: list[dict[str, Any]] = []
for index, item in enumerate(payload):
if not isinstance(item, dict):
raise ValidationError(f"benchmark row {index} must be an object")
rows.append(dict(item))
return rows
def _enrich_result(row: dict[str, Any], eval_entries: list[ManifestEntry]) -> dict[str, Any]:
enriched = dict(row)
entry = _entry_for_input(str(row.get("input_path", "")), eval_entries)
if entry:
enriched["sample_id"] = row.get("sample_id") or entry.sample_id
existing_slices = row.get("slices") if isinstance(row.get("slices"), list) else []
enriched["slices"] = sorted({*existing_slices, *_slices_for_entry(entry)})
else:
enriched.setdefault("slices", [])
return enriched
def _entry_for_input(input_path: str, entries: list[ManifestEntry]) -> ManifestEntry | None:
if not entries:
return None
path = Path(input_path)
try:
resolved = path.resolve()
except OSError:
resolved = path
for entry in entries:
entry_path = Path(entry.image_path)
try:
resolved_entry = entry_path.resolve()
except OSError:
resolved_entry = entry_path
if resolved == resolved_entry or input_path == entry.image_path:
return entry
return None
def _slices_for_entry(entry: ManifestEntry) -> list[str]:
values: set[str] = set()
metadata = entry.metadata or {}
for key in ("slice", "script", "document_type", "degradation", "language"):
value = metadata.get(key)
if isinstance(value, str) and value:
values.add(value)
for key in ("slices", "scripts", "tags"):
value = metadata.get(key)
if isinstance(value, list):
values.update(str(item) for item in value if str(item))
elif isinstance(value, str) and value:
values.add(value)
return sorted(values)
def _slice_counts(rows: list[dict[str, Any]]) -> dict[str, int]:
counts: dict[str, int] = {}
seen: set[tuple[str, str]] = set()
for row in rows:
key = sample_key(row)
for slice_name in row.get("slices", []) or []:
pair = (key, str(slice_name))
if pair in seen:
continue
seen.add(pair)
counts[str(slice_name)] = counts.get(str(slice_name), 0) + 1
return counts
def _full_eval_manifest_coverage_failures(
rows: list[dict[str, Any]],
config: ValidationConfig,
eval_entries: list[ManifestEntry],
) -> list[str]:
if not config.require_full_eval_manifest_coverage:
return []
if not eval_entries:
return ["eval manifest is required when require_full_eval_manifest_coverage is true"]
eval_sample_ids = {entry.sample_id for entry in eval_entries}
required_baselines = {config.candidate, *(gate.baseline for gate in config.gates)}
failures: list[str] = []
for baseline in sorted(required_baselines):
ok_sample_ids = {
str(row.get("sample_id"))
for row in rows
if row.get("baseline") == baseline and row.get("status") == "ok" and str(row.get("sample_id") or "")
}
missing = sorted(eval_sample_ids - ok_sample_ids)
if missing:
preview = ", ".join(missing[:5])
suffix = "" if len(missing) <= 5 else f", ... (+{len(missing) - 5} more)"
failures.append(
f"baseline {baseline!r} is missing successful rows for {len(missing)} eval manifest samples: {preview}{suffix}"
)
return failures
def _evaluate_gate(rows: list[dict[str, Any]], config: ValidationConfig, gate: ClaimGate) -> GateResult:
gate_slices = gate.slices or ([gate.slice] if gate.slice else [])
candidate = _rows_by_sample(rows, config.candidate, gate.metric, gate_slices)
baseline = _rows_by_sample(rows, gate.baseline, gate.metric, gate_slices)
shared = sorted(set(candidate) & set(baseline))
if not shared:
return GateResult(gate, False, 0, None, None, None, None, None, None, 0, 0, 0, None, None, None, f"no paired samples for {config.candidate} vs {gate.baseline} on {gate.metric}")
candidate_values = [candidate[key] for key in shared]
baseline_values = [baseline[key] for key in shared]
improvements = [
_absolute_improvement(candidate_value, baseline_value, gate.direction)
for candidate_value, baseline_value in zip(candidate_values, baseline_values, strict=True)
]
candidate_mean = sum(candidate_values) / len(candidate_values)
baseline_mean = sum(baseline_values) / len(baseline_values)
absolute = _absolute_improvement(candidate_mean, baseline_mean, gate.direction)
relative = _relative_improvement(absolute, baseline_mean)
ci_low, ci_high = _bootstrap_ci(improvements, config.bootstrap_iterations, config.bootstrap_confidence, config.random_seed)
wins, losses, ties = _paired_outcome_counts(candidate_values, baseline_values, gate.direction)
non_ties = wins + losses
win_rate = (wins / non_ties) if non_ties else None
pair_win_rate = wins / len(shared)
sign_test_p_value = _sign_test_p_value(wins, losses)
reasons: list[str] = []
if len(shared) < gate.min_pairs:
reasons.append(f"paired samples {len(shared)} < required {gate.min_pairs}")
if absolute < gate.min_absolute_margin:
reasons.append(f"absolute improvement {absolute:.6f} < required {gate.min_absolute_margin:.6f}")
if relative < gate.min_relative_improvement:
reasons.append(f"relative improvement {relative:.6f} < required {gate.min_relative_improvement:.6f}")
if gate.min_pair_win_rate is not None and pair_win_rate < gate.min_pair_win_rate:
reasons.append(f"paired win rate {pair_win_rate:.6f} < required {gate.min_pair_win_rate:.6f}")
if gate.max_p_value is not None:
if sign_test_p_value is None:
reasons.append("sign test p-value is unavailable because all paired outcomes are ties")
elif sign_test_p_value > gate.max_p_value:
reasons.append(f"sign test p-value {sign_test_p_value:.6f} > required {gate.max_p_value:.6f}")
if gate.require_ci_above_zero and ci_low <= 0:
reasons.append(f"bootstrap CI low {ci_low:.6f} <= 0")
return GateResult(
gate=gate,
passed=not reasons,
pairs=len(shared),
candidate_mean=candidate_mean,
baseline_mean=baseline_mean,
absolute_improvement=absolute,
relative_improvement=relative,
improvement_ci_low=ci_low,
improvement_ci_high=ci_high,
win_count=wins,
loss_count=losses,
tie_count=ties,
win_rate=win_rate,
pair_win_rate=pair_win_rate,
sign_test_p_value=sign_test_p_value,
reason="; ".join(reasons) if reasons else None,
)
def gate_has_quantitative_evidence(item: dict[str, Any]) -> bool:
gate = item.get("gate")
gate_payload = gate if isinstance(gate, dict) else {}
pairs = item.get("pairs")
if isinstance(pairs, bool) or not isinstance(pairs, int | float) or int(pairs) <= 0:
return False
pair_count = int(pairs)
for key in ("candidate_mean", "baseline_mean", "absolute_improvement"):
value = item.get(key)
if isinstance(value, bool) or not isinstance(value, int | float) or not math.isfinite(float(value)):
return False
counts: list[int] = []
for key in ("win_count", "loss_count", "tie_count"):
value = item.get(key)
if isinstance(value, bool) or not isinstance(value, int | float):
return False
count = int(value)
if count < 0:
return False
counts.append(count)
if sum(counts) != pair_count:
return False
min_pairs = gate_payload.get("min_pairs")
if min_pairs is not None:
if isinstance(min_pairs, bool) or not isinstance(min_pairs, int | float):
return False
if pair_count < int(min_pairs):
return False
min_absolute_margin = gate_payload.get("min_absolute_margin")
if min_absolute_margin is not None and not _numeric_threshold_passes(item.get("absolute_improvement"), min_absolute_margin):
return False
min_relative_improvement = gate_payload.get("min_relative_improvement")
if min_relative_improvement is not None and not _numeric_threshold_passes(item.get("relative_improvement"), min_relative_improvement):
return False
min_pair_win_rate = gate_payload.get("min_pair_win_rate")
if min_pair_win_rate is not None and not _numeric_threshold_passes(item.get("pair_win_rate"), min_pair_win_rate):
return False
max_p_value = gate_payload.get("max_p_value")
if max_p_value is not None and not _numeric_max_threshold_passes(item.get("sign_test_p_value"), max_p_value):
return False
if bool(gate_payload.get("require_ci_above_zero", True)):
ci_low = item.get("improvement_ci_low")
if isinstance(ci_low, bool) or not isinstance(ci_low, int | float) or not math.isfinite(float(ci_low)) or float(ci_low) <= 0:
return False
return True
def _numeric_threshold_passes(value: Any, threshold: Any) -> bool:
if isinstance(value, bool) or isinstance(threshold, bool):
return False
if not isinstance(value, int | float) or not isinstance(threshold, int | float):
return False
if not math.isfinite(float(value)) or not math.isfinite(float(threshold)):
return False
return float(value) >= float(threshold)
def _numeric_max_threshold_passes(value: Any, threshold: Any) -> bool:
if isinstance(value, bool) or isinstance(threshold, bool):
return False
if not isinstance(value, int | float) or not isinstance(threshold, int | float):
return False
if not math.isfinite(float(value)) or not math.isfinite(float(threshold)):
return False
return float(value) <= float(threshold)
def _rows_by_sample(rows: list[dict[str, Any]], baseline: str, metric: str, slice_names: list[str] | None) -> dict[str, float]:
values: dict[str, float] = {}
required_slices = {str(item) for item in slice_names or [] if str(item)}
for row in rows:
if row.get("baseline") != baseline or row.get("status") != "ok":
continue
row_slices = {str(item) for item in row.get("slices") or []}
if required_slices and not required_slices.issubset(row_slices):
continue
metrics = row.get("metrics") or {}
if metric not in metrics:
continue
value = metrics[metric]
if isinstance(value, int | float) and math.isfinite(float(value)):
values[sample_key(row)] = float(value)
return values
def _absolute_improvement(candidate_value: float, baseline_value: float, direction: str) -> float:
if direction == "lower":
return baseline_value - candidate_value
return candidate_value - baseline_value
def _relative_improvement(absolute_improvement: float, baseline_mean: float) -> float:
denominator = abs(baseline_mean)
if denominator == 0:
return math.inf if absolute_improvement > 0 else 0.0
return absolute_improvement / denominator
def _bootstrap_ci(values: list[float], iterations: int, confidence: float, seed: int) -> tuple[float, float]:
if not values:
return 0.0, 0.0
if len(values) == 1 or iterations <= 0:
mean = sum(values) / len(values)
return mean, mean
rng = random.Random(seed)
means: list[float] = []
for _ in range(iterations):
sample = [values[rng.randrange(len(values))] for _ in values]
means.append(sum(sample) / len(sample))
means.sort()
alpha = 1 - confidence
low_index = max(0, min(len(means) - 1, int((alpha / 2) * len(means))))
high_index = max(0, min(len(means) - 1, int((1 - alpha / 2) * len(means)) - 1))
return means[low_index], means[high_index]
def _paired_outcome_counts(candidate_values: list[float], baseline_values: list[float], direction: str) -> tuple[int, int, int]:
wins = 0
losses = 0
ties = 0
for candidate_value, baseline_value in zip(candidate_values, baseline_values, strict=True):
improvement = _absolute_improvement(candidate_value, baseline_value, direction)
if improvement > 0:
wins += 1
elif improvement < 0:
losses += 1
else:
ties += 1
return wins, losses, ties
def _sign_test_p_value(wins: int, losses: int) -> float | None:
trials = wins + losses
if trials == 0:
return None
numerator = sum(math.comb(trials, success_count) for success_count in range(wins, trials + 1))
return numerator / (2**trials)
def _aggregate_metrics(rows: list[dict[str, Any]]) -> dict[str, dict[str, dict[str, float | int]]]:
grouped: dict[str, dict[str, dict[str, list[float]]]] = {}
for row in rows:
if row.get("status") != "ok":
continue
baseline = str(row.get("baseline") or "")
if not baseline:
continue
slices = [*({str(item) for item in row.get("slices") or []} or {"all"})]
slices.append("all")
metrics = row.get("metrics")
if not isinstance(metrics, dict):
continue
for slice_name in slices:
slice_bucket = grouped.setdefault(baseline, {}).setdefault(slice_name, {})
for metric_name, value in metrics.items():
if isinstance(value, int | float) and math.isfinite(float(value)):
slice_bucket.setdefault(str(metric_name), []).append(float(value))
aggregates: dict[str, dict[str, dict[str, float | int]]] = {}
for baseline, slice_map in grouped.items():
aggregates[baseline] = {}
for slice_name, metric_map in slice_map.items():
aggregates[baseline][slice_name] = {
metric_name: sum(values) / len(values)
for metric_name, values in sorted(metric_map.items())
if values
}
if metric_map:
first_values = next(iter(metric_map.values()))
aggregates[baseline][slice_name]["sample_count"] = len(first_values)
return aggregates
def _hash_overlap(train_entries: list[ManifestEntry], eval_entries: list[ManifestEntry], attr: str) -> list[str]:
train_values = {getattr(entry, attr) for entry in train_entries if getattr(entry, attr)}
eval_values = {getattr(entry, attr) for entry in eval_entries if getattr(entry, attr)}
return sorted(str(item) for item in train_values & eval_values)
def _text_hash_overlap(train_entries: list[ManifestEntry], eval_entries: list[ManifestEntry]) -> list[str]:
def text_hash(entry: ManifestEntry) -> str:
value = entry.metadata.get("text_sha256") if entry.metadata else None
return str(value) if value else sha256_text(entry.text)
return sorted({text_hash(entry) for entry in train_entries} & {text_hash(entry) for entry in eval_entries})
def _sample_hash_overlap(train_entries: list[ManifestEntry], eval_entries: list[ManifestEntry]) -> list[str]:
def sample_hash(entry: ManifestEntry) -> str:
value = entry.metadata.get("sample_sha256") if entry.metadata else None
if value:
return str(value)
image_hash = entry.sha256 or ""
return sha256_text(f"{image_hash}\n{sha256_text(entry.text)}")
return sorted({sample_hash(entry) for entry in train_entries} & {sample_hash(entry) for entry in eval_entries})
def _write_validation_report(report: ValidationReport, output_dir: Path) -> None:
output_dir.mkdir(parents=True, exist_ok=True)
(output_dir / "validation.json").write_text(json.dumps(report.to_dict(), ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
lines = [
"# OCR Claim Validation",
"",
f"Status: `{report.claim_status}`",
f"Samples: `{report.sample_count}`",
f"Leakage check: `{'pass' if report.leakage.passed else 'fail'}`",
"",
"## Missing Requirements",
"",
]
if report.missing_requirements:
lines.extend(f"- {item}" for item in report.missing_requirements)
else:
lines.append("- none")
lines.extend(
[
"",
"## Gates",
"",
"| baseline | slice | metric | pairs | wins | losses | ties | win_rate | pair_win_rate | p_value | candidate | baseline_mean | improvement | ci_low | pass | reason |",
"| --- | --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | --- | --- |",
]
)
for gate in report.gates:
lines.append(
"| {baseline} | {slice} | {metric} | {pairs} | {wins} | {losses} | {ties} | {win_rate} | {pair_win_rate} | {p_value} | {candidate} | {baseline_mean} | {improvement} | {ci_low} | {passed} | {reason} |".format(
baseline=gate.gate.baseline,
slice=",".join(gate.gate.slices or ([gate.gate.slice] if gate.gate.slice else [])),
metric=gate.gate.metric,
pairs=gate.pairs,
wins=gate.win_count,
losses=gate.loss_count,
ties=gate.tie_count,
win_rate=f"{gate.win_rate:.4f}" if gate.win_rate is not None else "",
pair_win_rate=f"{gate.pair_win_rate:.4f}" if gate.pair_win_rate is not None else "",
p_value=f"{gate.sign_test_p_value:.6f}" if gate.sign_test_p_value is not None else "",
candidate=f"{gate.candidate_mean:.6f}" if gate.candidate_mean is not None else "",
baseline_mean=f"{gate.baseline_mean:.6f}" if gate.baseline_mean is not None else "",
improvement=f"{gate.absolute_improvement:.6f}" if gate.absolute_improvement is not None else "",
ci_low=f"{gate.improvement_ci_low:.6f}" if gate.improvement_ci_low is not None else "",
passed="yes" if gate.passed else "no",
reason=(gate.reason or "").replace("|", "\\|"),
)
)
lines.extend(["", "## Aggregate Metrics", ""])
for baseline, slice_map in sorted(report.aggregate_metrics.items()):
lines.append(f"### {baseline}")
lines.append("")
lines.append("| slice | sample_count | metric | mean |")
lines.append("| --- | ---: | --- | ---: |")
for slice_name, metrics in sorted(slice_map.items()):
sample_count = metrics.get("sample_count", "")
for metric_name, value in sorted(metrics.items()):
if metric_name == "sample_count":
continue
lines.append(f"| {slice_name} | {sample_count} | {metric_name} | {float(value):.6f} |")
lines.append("")
(output_dir / "validation.md").write_text("\n".join(lines) + "\n", encoding="utf-8")
|