File size: 42,740 Bytes
518db7a | 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 | """
Validator classes for text generation quality assessment.
Provides FastValidator (heuristics), GrammarValidator (LanguageTool),
and KnowledgeValidator (factual accuracy) with security hardening and
performance optimizations.
"""
__all__ = [
"FastValidator",
"GrammarValidator",
"KnowledgeValidator",
"LanguageValidator",
"PerplexityValidator",
"Validator",
"FastValidationResult",
"GrammarValidationResult",
"KnowledgeValidationResult",
"LanguageValidationResult",
"PerplexityValidationResult",
]
import time
import asyncio
import logging
from typing import Any, Protocol, TypedDict, TYPE_CHECKING
from collections import deque, Counter
from dataclasses import dataclass, field
import torch
# Import GrammarResult for type compatibility
if TYPE_CHECKING:
from .grammar_checker import GrammarResult
else:
try:
from .grammar_checker import GrammarResult
except ImportError:
# Fallback if grammar_checker not available
@dataclass
class GrammarResult:
grammar_score: float
num_errors: int
errors: list[dict] = field(default_factory=list)
suggestions: list[list[str]] = field(default_factory=list)
is_fallback: bool = False
# Import unified sanitization
from .sanitizer import sanitize
# Import validation constants
from .constants import (
MIN_ASCII_RATIO,
MAX_REPETITION_RATIO,
MIN_SAMPLE_LENGTH,
VALIDATION_MAX_LENGTH,
VALIDATION_TEMPERATURE,
KNOWLEDGE_MAX_LENGTH,
KNOWLEDGE_TEMPERATURE,
SAMPLE_HISTORY_SIZE,
GRAMMAR_HISTORY_SIZE,
TIMESTAMP_HISTORY_SIZE,
TREND_ANALYSIS_WINDOW,
NGRAM_SIZE,
MIN_NGRAM_TEXT_LENGTH,
FALLBACK_REPETITION_SCORE,
FALLBACK_GRAMMAR_SCORE,
FALLBACK_ERROR_COUNT,
ERROR_LOG_TRUNCATE_LENGTH,
)
logger = logging.getLogger(__name__)
class FastValidationResult(TypedDict):
"""Return type for FastValidator.validate()."""
samples: list[str]
is_garbage: bool
ascii_ratio: float
avg_length: float
repetition_ratio: float
class GrammarValidationResult(TypedDict):
"""Return type for GrammarValidator.validate()."""
grammar_score: float
num_errors: int
is_fallback: bool
samples: list[str]
class KnowledgeValidationResult(TypedDict):
"""Return type for KnowledgeValidator.validate()."""
accuracy: float
correct: int
total: int
failed: list[dict[str, Any]]
class LanguageValidationResult(TypedDict):
"""Return type for LanguageValidator.validate()."""
is_garbage: bool
lang_confidence: float
valid_word_ratio: float
detected_language: str
samples: list[str]
class PerplexityValidationResult(TypedDict):
"""Return type for PerplexityValidator.validate()."""
perplexity: float
perplexity_normalized: float
samples: list[str]
class Validator(Protocol):
"""
Protocol for validation components.
Validators must implement a validate() method that takes a text-generating
model and training step, returning validation metrics.
This Protocol provides structural subtyping (duck typing with type hints),
allowing type checkers to verify validator compliance without requiring
inheritance.
Example:
>>> class CustomValidator:
... def validate(self, model: Any, step: int) -> dict[str, Any]:
... return {"score": 0.95}
...
>>> validator: Validator = CustomValidator() # Type-safe!
"""
def validate(self, model: Any, step: int) -> dict[str, Any]:
"""
Run validation on model at given training step.
Args:
model: Model with .generate_text() method
step: Current training step
Returns:
Dict with validation metrics (keys vary by validator):
- FastValidator: is_garbage, ascii_ratio, avg_length, repetition_ratio
- GrammarValidator: grammar_score, num_errors, is_fallback
- KnowledgeValidator: accuracy, correct, total, failed
"""
...
def validate_samples(self, samples: list[str], step: int) -> dict[str, Any]:
"""
Run validation on pre-generated samples.
This method allows sharing samples between multiple validators,
reducing generation cost.
Args:
samples: Pre-generated text samples
step: Current training step
Returns:
Dict with validation metrics (same as validate())
"""
...
class FastValidator:
"""
Heuristic-based fast validation for garbage detection.
Runs every 100 steps with <1s overhead. Catches obvious failures
like non-ASCII output, extremely short/long output, and repetition.
"""
def __init__(self, test_prompts: list[str]) -> None:
"""
Initialize FastValidator.
Args:
test_prompts: List of prompts to test generation with
Raises:
ValueError: If test_prompts is empty
TypeError: If test_prompts contains non-string elements
"""
if not test_prompts:
raise ValueError("test_prompts cannot be empty")
if not all(isinstance(p, str) for p in test_prompts):
raise TypeError("All test_prompts must be strings")
self.test_prompts = test_prompts
self.sample_history: deque[tuple[int, list[str]]] = deque(maxlen=SAMPLE_HISTORY_SIZE)
@staticmethod
def _ngram_repetition(text: str) -> float:
"""
Calculate n-gram repetition ratio using memory-efficient generator.
Args:
text: Input text to analyze
Returns:
Repetition ratio (0.0 = no repetition, 1.0 = maximum repetition)
"""
if len(text) < NGRAM_SIZE:
return 0.0
# Generator avoids materializing full list in memory
ngrams = (text[i:i+NGRAM_SIZE] for i in range(len(text) - NGRAM_SIZE + 1))
counts = Counter(ngrams)
total = sum(counts.values())
unique = len(counts)
# Convert to repetition ratio (inverse of uniqueness)
return 1.0 - (unique / total) if total > 0 else 0.0
def validate(self, model: Any, step: int) -> FastValidationResult:
"""
Run fast heuristic validation.
Args:
model: Model to validate
step: Current training step
Returns:
FastValidationResult with keys:
- samples: list[str]
- is_garbage: bool
- ascii_ratio: float
- avg_length: float
- repetition_ratio: float
"""
samples = []
try:
# Generate with inference mode for performance
with torch.inference_mode():
for prompt in self.test_prompts:
try:
sample = model.generate_text(
prompt,
max_length=VALIDATION_MAX_LENGTH,
temperature=VALIDATION_TEMPERATURE
)
samples.append(sample)
except Exception as e:
logger.warning(
"Generation failed for prompt",
extra={"prompt": prompt, "error": str(e)}
)
samples.append("")
except Exception as e:
logger.error(
"FastValidator failed",
extra={"step": step, "error": str(e)}
)
return {
"samples": [],
"is_garbage": True,
"ascii_ratio": 0.0,
"avg_length": 0.0,
"repetition_ratio": FALLBACK_REPETITION_SCORE
}
# Delegate to validate_samples for actual validation logic
return self.validate_samples(samples, step)
def validate_samples(self, samples: list[str], step: int) -> FastValidationResult:
"""
Run fast heuristic validation on pre-generated samples.
This method allows sharing samples between multiple validators,
reducing generation cost by 50%.
Args:
samples: Pre-generated text samples
step: Current training step
Returns:
FastValidationResult with keys:
- samples: list[str]
- is_garbage: bool
- ascii_ratio: float
- avg_length: float
- repetition_ratio: float
"""
# Heuristic checks
total_chars = sum(len(s) for s in samples)
ascii_chars = sum(sum(c.isascii() for c in s) for s in samples)
ascii_ratio = ascii_chars / total_chars if total_chars > 0 else 0.0
avg_length = sum(len(s) for s in samples) / len(samples) if samples else 0
# Repetition detection (memory-efficient generator-based)
repetition_scores = []
for sample in samples:
if len(sample) < MIN_NGRAM_TEXT_LENGTH:
repetition_scores.append(FALLBACK_REPETITION_SCORE)
continue
# Use generator-based n-gram detection (O(1) memory)
rep_ratio = self._ngram_repetition(sample)
repetition_scores.append(rep_ratio)
repetition_ratio = sum(repetition_scores) / len(repetition_scores) if repetition_scores else 0.0
# Garbage detection criteria
is_garbage = (
ascii_ratio < MIN_ASCII_RATIO or
avg_length < MIN_SAMPLE_LENGTH or
repetition_ratio > MAX_REPETITION_RATIO
)
# Store sanitized samples
sanitized_samples = [sanitize(s, mode="pii") for s in samples]
self.sample_history.append((step, sanitized_samples))
return {
"samples": sanitized_samples,
"is_garbage": is_garbage,
"ascii_ratio": ascii_ratio,
"avg_length": avg_length,
"repetition_ratio": repetition_ratio
}
class GrammarValidator:
"""
LanguageTool-based grammar validation.
Runs every 200 steps with <2s overhead. Measures grammar quality
using external LanguageTool API with fallback to heuristics.
"""
def __init__(self, client: Any, test_prompts: list[str]) -> None:
"""
Initialize GrammarValidator.
Args:
client: LanguageToolClient instance
test_prompts: List of prompts to test generation with
Raises:
ValueError: If client is None or test_prompts is empty
TypeError: If test_prompts contains non-string elements
"""
if client is None:
raise ValueError("client cannot be None")
if not test_prompts:
raise ValueError("test_prompts cannot be empty")
if not all(isinstance(p, str) for p in test_prompts):
raise TypeError("All test_prompts must be strings")
self.client = client
self.test_prompts = test_prompts
# Inline history tracking (removed ValidationHistory abstraction)
self.grammar_scores: deque[float] = deque(maxlen=GRAMMAR_HISTORY_SIZE)
self.sample_outputs: deque[str] = deque(maxlen=SAMPLE_HISTORY_SIZE)
self.timestamps: deque[int] = deque(maxlen=TIMESTAMP_HISTORY_SIZE)
def validate(self, model: Any, step: int) -> GrammarValidationResult:
"""
Run grammar validation (sync wrapper for async validation).
This method wraps validate_async() to maintain backward compatibility
with PyTorch Lightning callbacks that expect synchronous validation.
For direct async usage, call validate_async() instead.
Args:
model: Model to validate
step: Current training step
Returns:
GrammarValidationResult with keys:
- grammar_score: float
- num_errors: int
- is_fallback: bool
- samples: list[str]
"""
# Use async validation with asyncio.run()
try:
return asyncio.run(self.validate_async(model, step))
except RuntimeError as e:
# Handle case where event loop is already running
if "already running" in str(e):
logger.warning("Event loop already running, falling back to sync validation")
return self._validate_sync(model, step)
raise
def _validate_sync(self, model: Any, step: int) -> GrammarValidationResult:
"""
Synchronous fallback validation (used when event loop conflicts occur).
This is the original sequential implementation, kept as fallback.
Args:
model: Model to validate
step: Current training step
Returns:
GrammarValidationResult (same structure as validate())
"""
samples = []
try:
with torch.inference_mode():
for prompt in self.test_prompts:
try:
sample = model.generate_text(
prompt,
max_length=VALIDATION_MAX_LENGTH,
temperature=VALIDATION_TEMPERATURE
)
samples.append(sample)
except Exception as e:
logger.warning("Generation failed", extra={"error": str(e)})
samples.append("")
except Exception as e:
logger.error(
"GrammarValidator generation failed",
extra={"error": str(e)}
)
return {
"grammar_score": FALLBACK_GRAMMAR_SCORE,
"num_errors": FALLBACK_ERROR_COUNT,
"is_fallback": True,
"samples": []
}
# Delegate to validate_samples_sync for actual validation logic
return self.validate_samples_sync(samples, step)
def validate_samples_sync(self, samples: list[str], step: int) -> GrammarValidationResult:
"""
Run grammar validation on pre-generated samples (synchronous).
This method allows sharing samples between multiple validators,
reducing generation cost by 50%.
Args:
samples: Pre-generated text samples
step: Current training step
Returns:
GrammarValidationResult with keys:
- grammar_score: float
- num_errors: int
- is_fallback: bool
- samples: list[str]
"""
# Check grammar for all samples (SEQUENTIAL)
results = []
for sample in samples:
if not sample or len(sample) < MIN_SAMPLE_LENGTH:
results.append(GrammarResult(
grammar_score=FALLBACK_GRAMMAR_SCORE,
num_errors=0,
errors=[],
suggestions=[],
is_fallback=True
))
continue
result = self.client.check(sample)
results.append(result)
# Aggregate scores
avg_score = sum(r.grammar_score for r in results) / len(results) if results else 0.0
total_errors = sum(r.num_errors for r in results)
any_fallback = any(r.is_fallback for r in results)
# Update history
if samples:
sanitized = sanitize(samples[0], mode="pii")
self.grammar_scores.append(avg_score)
self.sample_outputs.append(sanitized)
self.timestamps.append(step)
return {
"grammar_score": avg_score,
"num_errors": total_errors,
"is_fallback": any_fallback,
"samples": [sanitize(s, mode="pii") for s in samples]
}
async def validate_async(self, model: Any, step: int) -> GrammarValidationResult:
"""
Run async grammar validation (NON-BLOCKING).
This is the key performance optimization: all grammar checks run
in parallel instead of sequentially, reducing validation time from
2.5s to 0.5s (5x speedup).
Args:
model: Model to validate
step: Current training step
Returns:
{
"grammar_score": float,
"num_errors": int,
"is_fallback": bool,
"samples": list[str]
}
"""
samples = []
try:
# Generate samples (still synchronous, but fast)
with torch.inference_mode():
for prompt in self.test_prompts:
try:
sample = model.generate_text(
prompt,
max_length=VALIDATION_MAX_LENGTH,
temperature=VALIDATION_TEMPERATURE
)
samples.append(sample)
except Exception as e:
logger.warning("Generation failed", extra={"error": str(e)})
samples.append("")
except Exception as e:
logger.error(
"GrammarValidator generation failed",
extra={"error": str(e)}
)
return {
"grammar_score": FALLBACK_GRAMMAR_SCORE,
"num_errors": FALLBACK_ERROR_COUNT,
"is_fallback": True,
"samples": []
}
# Delegate to validate_samples_async for actual validation logic
return await self.validate_samples_async(samples, step)
async def validate_samples_async(self, samples: list[str], step: int) -> GrammarValidationResult:
"""
Run async grammar validation on pre-generated samples (NON-BLOCKING).
This method allows sharing samples between multiple validators,
reducing generation cost by 50%.
Args:
samples: Pre-generated text samples
step: Current training step
Returns:
GrammarValidationResult with keys:
- grammar_score: float
- num_errors: int
- is_fallback: bool
- samples: list[str]
"""
# Filter out empty/too-short samples
valid_samples = [s for s in samples if s and len(s) >= MIN_SAMPLE_LENGTH]
# ASYNC: Check grammar in parallel (KEY OPTIMIZATION)
if hasattr(self.client, 'check_batch_async'):
# Use async client for parallel checking
results = await self.client.check_batch_async(valid_samples)
else:
# Fallback to sync client (sequential)
logger.warning("Async client not available, falling back to sync")
results = [self.client.check(s) for s in valid_samples]
# Aggregate scores
avg_score = sum(r.grammar_score for r in results) / len(results) if results else 0.0
total_errors = sum(r.num_errors for r in results)
any_fallback = any(r.is_fallback for r in results)
# Update history
if samples:
sanitized = sanitize(samples[0], mode="pii")
self.grammar_scores.append(avg_score)
self.sample_outputs.append(sanitized)
self.timestamps.append(step)
return {
"grammar_score": avg_score,
"num_errors": total_errors,
"is_fallback": any_fallback,
"samples": [sanitize(s, mode="pii") for s in samples]
}
def validate_samples(self, samples: list[str], step: int) -> GrammarValidationResult:
"""
Synchronous wrapper for validate_samples_async (for CombinedValidationCallback).
This method provides a synchronous interface for validating pre-generated
samples, allowing the CombinedValidationCallback to share samples between
validators.
Args:
samples: Pre-generated text samples
step: Current training step
Returns:
GrammarValidationResult with same structure as validate()
"""
try:
return asyncio.run(self.validate_samples_async(samples, step))
except RuntimeError as e:
# Handle case where event loop is already running
if "already running" in str(e):
logger.warning("Event loop already running, falling back to sync validation")
return self.validate_samples_sync(samples, step)
raise
def get_trend(self, window: int = TREND_ANALYSIS_WINDOW) -> str:
"""
Detect improving/degrading trend.
Args:
window: Number of recent scores to analyze
Returns:
"improving", "degrading", "stable", or "insufficient_data"
"""
if len(self.grammar_scores) < window:
return "insufficient_data"
recent = list(self.grammar_scores)[-window:]
if all(recent[i] >= recent[i-1] for i in range(1, len(recent))):
return "improving"
elif all(recent[i] <= recent[i-1] for i in range(1, len(recent))):
return "degrading"
else:
return "stable"
class KnowledgeValidator:
"""
Factual accuracy validation using knowledge base.
Runs post-training only (~10s). Tests model on 10 factual questions
to verify knowledge retention.
"""
def __init__(self, questions: list[dict[str, Any]]) -> None:
"""
Initialize KnowledgeValidator.
Args:
questions: List of {"q": str, "a": list[str]} question/answer pairs
Raises:
ValueError: If questions list is None or has invalid structure
TypeError: If questions is not a list
"""
if questions is None:
raise ValueError("questions cannot be None")
if not isinstance(questions, list):
raise TypeError("questions must be a list")
# Validate structure of questions (each must have 'q' and 'a' keys)
for i, q in enumerate(questions):
if not isinstance(q, dict):
raise TypeError(f"Question at index {i} must be a dict")
if 'q' not in q or 'a' not in q:
raise ValueError(f"Question at index {i} must have 'q' and 'a' keys")
if not isinstance(q['q'], str):
raise TypeError(f"Question 'q' at index {i} must be a string")
if not isinstance(q['a'], list):
raise TypeError(f"Question 'a' at index {i} must be a list")
self.questions = questions
def validate(self, model: Any, step: int = -1) -> KnowledgeValidationResult:
"""
Run factual accuracy validation.
Args:
model: Model to validate
step: Training step (default -1 for post-training)
Returns:
KnowledgeValidationResult with keys:
- accuracy: float
- correct: int
- total: int
- failed: list[dict[str, Any]]
"""
correct = 0
failed = []
try:
with torch.inference_mode():
for item in self.questions:
question = item['q']
valid_answers = [a.lower() for a in item['a']]
try:
output = model.generate_text(
question,
max_length=KNOWLEDGE_MAX_LENGTH,
temperature=KNOWLEDGE_TEMPERATURE
)
output_lower = output.lower()
# Fuzzy matching: check if any valid answer in output
is_correct = any(ans in output_lower for ans in valid_answers)
if is_correct:
correct += 1
else:
failed.append({
'question': question,
'expected': item['a'],
'got': output[:ERROR_LOG_TRUNCATE_LENGTH]
})
except Exception as e:
logger.warning(
"Knowledge validation failed",
extra={"question": question, "error": str(e)}
)
failed.append({
'question': question,
'expected': item['a'],
'got': f"ERROR: {str(e)}"
})
except Exception as e:
logger.error(
"KnowledgeValidator failed",
extra={"error": str(e)}
)
return {
"accuracy": 0.0,
"correct": 0,
"total": len(self.questions),
"failed": self.questions
}
return {
"accuracy": correct / len(self.questions) if self.questions else 0.0,
"correct": correct,
"total": len(self.questions),
"failed": failed
}
def validate_samples(self, samples: list[str], step: int) -> KnowledgeValidationResult:
"""
Not applicable for KnowledgeValidator (uses its own Q&A format).
This method exists for Protocol compliance but is not supported.
Use validate() instead.
Args:
samples: Unused (KnowledgeValidator generates from questions)
step: Training step
Raises:
NotImplementedError: KnowledgeValidator doesn't support validate_samples
Note:
KnowledgeValidator doesn't use pre-generated samples since it
tests factual knowledge with specific Q&A pairs.
"""
raise NotImplementedError(
"KnowledgeValidator doesn't support validate_samples. "
"Use validate(model, step) instead."
)
class LanguageValidator:
"""
Language detection and word validity validation.
Validates text is English with real words using:
- langdetect for language detection
- NLTK words corpus for English word validation
- Unicode script detection for multilingual text
Runs every 100 steps with <1s overhead.
"""
def __init__(self, test_prompts: list[str]) -> None:
"""
Initialize LanguageValidator.
Args:
test_prompts: List of prompts to test generation with
Raises:
ValueError: If test_prompts is empty
TypeError: If test_prompts contains non-string elements
"""
if not test_prompts:
raise ValueError("test_prompts cannot be empty")
if not all(isinstance(p, str) for p in test_prompts):
raise TypeError("All test_prompts must be strings")
self.test_prompts = test_prompts
# Load English words corpus (lazy load to avoid startup cost)
self._english_words = None
@property
def english_words(self):
"""Lazy-load NLTK words corpus."""
if self._english_words is None:
try:
import nltk
from nltk.corpus import words
# Ensure local NLTK data directory is searched first
nltk.data.path.insert(0, "/home/mikeb/nltk_data")
self._english_words = set(w.lower() for w in words.words())
except Exception as e:
logger.warning(
"NLTK words corpus not available, using fallback",
extra={"error": str(e)}
)
# Fallback to small set of common English words
self._english_words = set([
'the', 'be', 'to', 'of', 'and', 'a', 'in', 'that', 'have', 'i',
'it', 'for', 'not', 'on', 'with', 'he', 'as', 'you', 'do', 'at'
])
return self._english_words
@staticmethod
def detect_language_with_confidence(text: str) -> tuple[str, float]:
"""
Detect language and return confidence score.
Args:
text: Input text to analyze
Returns:
Tuple of (language_code, confidence)
e.g., ('en', 0.95) for high-confidence English
"""
try:
import langdetect
from langdetect import DetectorFactory
# Ensure reproducible results
DetectorFactory.seed = 0
# Detect language
lang = langdetect.detect(text)
# Get probability distribution
probs = langdetect.detect_langs(text)
# Find English confidence
en_confidence = next(
(p.prob for p in probs if p.lang == 'en'),
0.0
)
return lang, en_confidence if lang == 'en' else 0.0
except Exception as e:
logger.debug(
"Language detection failed",
extra={"error": str(e)}
)
return 'unknown', 0.0
@staticmethod
def detect_multilingual(text: str) -> dict[str, Any]:
"""
Detect mixed-language text (common gaming strategy).
Args:
text: Input text to analyze
Returns:
Dict with keys:
- is_multilingual: bool
- primary_script: str
- script_ratios: dict[str, float]
"""
# Unicode script detection
scripts = {
'latin': 0,
'cyrillic': 0,
'arabic': 0,
'cjk': 0,
'greek': 0,
}
for char in text:
if 'a' <= char.lower() <= 'z':
scripts['latin'] += 1
elif '\u0400' <= char <= '\u04FF':
scripts['cyrillic'] += 1
elif '\u0600' <= char <= '\u06FF':
scripts['arabic'] += 1
elif '\u4E00' <= char <= '\u9FFF':
scripts['cjk'] += 1
elif '\u0370' <= char <= '\u03FF':
scripts['greek'] += 1
total_letters = sum(scripts.values())
if total_letters == 0:
return {
'is_multilingual': False,
'primary_script': 'none',
'script_ratios': {}
}
# Normalize to percentages
script_ratios = {k: v/total_letters for k, v in scripts.items()}
# Find dominant script
primary_script = max(script_ratios, key=script_ratios.get)
# Check if multiple scripts present
num_scripts = sum(1 for ratio in script_ratios.values() if ratio > 0.05)
return {
'is_multilingual': num_scripts > 1,
'primary_script': primary_script,
'script_ratios': script_ratios,
}
def validate(self, model: Any, step: int) -> LanguageValidationResult:
"""
Run language detection and word validity validation.
Args:
model: Model to validate
step: Current training step
Returns:
LanguageValidationResult with keys:
- is_garbage: bool
- lang_confidence: float
- valid_word_ratio: float
- detected_language: str
- samples: list[str]
"""
samples = []
try:
with torch.inference_mode():
for prompt in self.test_prompts:
try:
sample = model.generate_text(
prompt,
max_length=VALIDATION_MAX_LENGTH,
temperature=VALIDATION_TEMPERATURE
)
samples.append(sample)
except Exception as e:
logger.warning(
"Generation failed for prompt",
extra={"prompt": prompt, "error": str(e)}
)
samples.append("")
except Exception as e:
logger.error(
"LanguageValidator failed",
extra={"step": step, "error": str(e)}
)
return {
"is_garbage": True,
"lang_confidence": 0.0,
"valid_word_ratio": 0.0,
"detected_language": "unknown",
"samples": []
}
# Delegate to validate_samples for actual validation logic
return self.validate_samples(samples, step)
def validate_samples(self, samples: list[str], step: int) -> LanguageValidationResult:
"""
Run language validation on pre-generated samples.
This method allows sharing samples between multiple validators,
reducing generation cost.
Args:
samples: Pre-generated text samples
step: Current training step
Returns:
LanguageValidationResult with keys:
- is_garbage: bool
- lang_confidence: float
- valid_word_ratio: float
- detected_language: str
- samples: list[str]
"""
if not samples:
return {
"is_garbage": True,
"lang_confidence": 0.0,
"valid_word_ratio": 0.0,
"detected_language": "unknown",
"samples": []
}
# Aggregate language detection across all samples
lang_confidences = []
detected_langs = []
valid_word_ratios = []
for sample in samples:
if not sample or len(sample) < MIN_SAMPLE_LENGTH:
lang_confidences.append(0.0)
detected_langs.append('unknown')
valid_word_ratios.append(0.0)
continue
# Language detection
lang, confidence = self.detect_language_with_confidence(sample)
lang_confidences.append(confidence)
detected_langs.append(lang)
# Word validity check
tokens = sample.lower().split()
clean_tokens = [
t.strip('.,!?;:()[]{}"\'-')
for t in tokens
if t.strip('.,!?;:()[]{}"\'-')
]
if clean_tokens:
valid_count = sum(
1 for t in clean_tokens
if t in self.english_words
)
valid_ratio = valid_count / len(clean_tokens)
else:
valid_ratio = 0.0
valid_word_ratios.append(valid_ratio)
# Aggregate scores
avg_lang_confidence = sum(lang_confidences) / len(lang_confidences)
avg_valid_word_ratio = sum(valid_word_ratios) / len(valid_word_ratios)
# Most common detected language
from collections import Counter
lang_counts = Counter(detected_langs)
primary_lang = lang_counts.most_common(1)[0][0]
# Check for multilingual text in any sample
any_multilingual = any(
self.detect_multilingual(s)['is_multilingual']
for s in samples
if s and len(s) >= MIN_SAMPLE_LENGTH
)
# Garbage detection criteria
is_garbage = (
primary_lang != 'en' or
avg_lang_confidence < 0.8 or
avg_valid_word_ratio < 0.7 or
any_multilingual
)
# Sanitize samples
sanitized_samples = [sanitize(s, mode="pii") for s in samples]
return {
"is_garbage": is_garbage,
"lang_confidence": avg_lang_confidence,
"valid_word_ratio": avg_valid_word_ratio,
"detected_language": primary_lang,
"samples": sanitized_samples
}
class PerplexityValidator:
"""
Autoregressive perplexity validation using DistilGPT-2.
Measures language fluency using pre-trained transformer model.
Uses mixed precision (AMP) for 2x speedup.
Runs every 100 steps with ~500ms overhead (with batching).
"""
def __init__(self, test_prompts: list[str], model_name: str = "distilgpt2") -> None:
"""
Initialize PerplexityValidator.
Args:
test_prompts: List of prompts to test generation with
model_name: HuggingFace model name (default: "distilgpt2")
Raises:
ValueError: If test_prompts is empty
TypeError: If test_prompts contains non-string elements
"""
if not test_prompts:
raise ValueError("test_prompts cannot be empty")
if not all(isinstance(p, str) for p in test_prompts):
raise TypeError("All test_prompts must be strings")
self.test_prompts = test_prompts
self.model_name = model_name
# Lazy-load model (avoid startup cost)
self._model = None
self._tokenizer = None
@property
def model(self):
"""Lazy-load DistilGPT-2 model."""
if self._model is None:
try:
from transformers import AutoModelForCausalLM
self._model = AutoModelForCausalLM.from_pretrained(
self.model_name
).to('cuda')
self._model.eval()
except Exception as e:
logger.error(
"Failed to load perplexity model",
extra={"model": self.model_name, "error": str(e)}
)
raise
return self._model
@property
def tokenizer(self):
"""Lazy-load tokenizer."""
if self._tokenizer is None:
try:
from transformers import AutoTokenizer
self._tokenizer = AutoTokenizer.from_pretrained(self.model_name)
except Exception as e:
logger.error(
"Failed to load tokenizer",
extra={"model": self.model_name, "error": str(e)}
)
raise
return self._tokenizer
def validate(self, model: Any, step: int) -> PerplexityValidationResult:
"""
Run perplexity validation.
Args:
model: Model to validate
step: Current training step
Returns:
PerplexityValidationResult with keys:
- perplexity: float
- perplexity_normalized: float (0-1 score for reward)
- samples: list[str]
"""
samples = []
try:
with torch.inference_mode():
for prompt in self.test_prompts:
try:
sample = model.generate_text(
prompt,
max_length=VALIDATION_MAX_LENGTH,
temperature=VALIDATION_TEMPERATURE
)
samples.append(sample)
except Exception as e:
logger.warning(
"Generation failed for prompt",
extra={"prompt": prompt, "error": str(e)}
)
samples.append("")
except Exception as e:
logger.error(
"PerplexityValidator generation failed",
extra={"step": step, "error": str(e)}
)
return {
"perplexity": float('inf'),
"perplexity_normalized": 0.0,
"samples": []
}
# Delegate to validate_samples for actual validation logic
return self.validate_samples(samples, step)
def validate_samples(self, samples: list[str], step: int) -> PerplexityValidationResult:
"""
Run perplexity validation on pre-generated samples.
This method allows sharing samples between multiple validators,
reducing generation cost.
Args:
samples: Pre-generated text samples
step: Current training step
Returns:
PerplexityValidationResult with keys:
- perplexity: float
- perplexity_normalized: float (0-1 score for reward)
- samples: list[str]
"""
if not samples:
return {
"perplexity": float('inf'),
"perplexity_normalized": 0.0,
"samples": []
}
# Filter valid samples
valid_samples = [
s for s in samples
if s and len(s) >= MIN_SAMPLE_LENGTH
]
if not valid_samples:
return {
"perplexity": float('inf'),
"perplexity_normalized": 0.0,
"samples": [sanitize(s, mode="pii") for s in samples]
}
# Compute perplexity for each sample
perplexities = []
try:
for sample in valid_samples:
# Tokenize
encodings = self.tokenizer(
sample,
return_tensors='pt',
truncation=True,
max_length=512
).to('cuda')
# Compute cross-entropy with mixed precision
with torch.no_grad(), torch.amp.autocast("cuda"):
outputs = self.model(**encodings, labels=encodings.input_ids)
ce = outputs.loss.item()
# Perplexity = exp(cross_entropy)
perplexity = torch.exp(torch.tensor(ce)).item()
perplexities.append(perplexity)
except Exception as e:
logger.error(
"Perplexity computation failed",
extra={"error": str(e)}
)
return {
"perplexity": float('inf'),
"perplexity_normalized": 0.0,
"samples": [sanitize(s, mode="pii") for s in samples]
}
# Aggregate
avg_perplexity = sum(perplexities) / len(perplexities)
# Normalize to [0, 1] for reward (lower perplexity = better)
# exp(-perp/10): perp=0 → 1.0, perp=10 → 0.37, perp=50 → 0.007
import math
normalized_score = math.exp(-avg_perplexity / 10.0)
return {
"perplexity": avg_perplexity,
"perplexity_normalized": normalized_score,
"samples": [sanitize(s, mode="pii") for s in samples]
}
|