File size: 21,895 Bytes
1a4aa87 | 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 | """
Benchmarking Engine
Main orchestration for the AegisLM Benchmarking Engine:
- Baseline evaluation mode
- Adversarial evaluation mode
- Delta robustness computation
- Cross-model comparison
- Statistical reporting
- Benchmark artifact generation
"""
import asyncio
import time
import uuid
from datetime import datetime
from enum import Enum
from typing import Any, Dict, List, Optional
from uuid import UUID
from backend.benchmarking.comparison import (
generate_comparative_report,
generate_vulnerability_heatmap,
rank_models,
)
from backend.benchmarking.reporter import (
generate_benchmark_artifact,
generate_text_report,
)
from backend.benchmarking.schemas import (
BenchmarkConfig,
BenchmarkMode,
BenchmarkPerformance,
BenchmarkResult,
BenchmarkStatus,
BenchmarkWeights,
EvaluationResult,
MetricDeltas,
ModelBenchmarkResult,
ModelMetrics,
StartBenchmarkRequest,
)
from backend.benchmarking.statistics import (
MetricStatistics,
calculate_vulnerability_consistency,
)
from backend.core.config import settings
from backend.core.orchestrator import (
EvaluationInput,
EvaluationOrchestrator,
RunStatus,
)
from backend.logging.logger import get_logger
# =============================================================================
# Benchmark Events
# =============================================================================
class BenchmarkEvent(str, Enum):
"""Observability events for benchmarking."""
BENCHMARK_STARTED = "BENCHMARK_STARTED"
BENCHMARK_COMPLETED = "BENCHMARK_COMPLETED"
BENCHMARK_FAILED = "BENCHMARK_FAILED"
MODEL_EVALUATION_STARTED = "MODEL_EVALUATION_STARTED"
MODEL_EVALUATION_COMPLETED = "MODEL_EVALUATION_COMPLETED"
BASELINE_COMPLETED = "BASELINE_COMPLETED"
ADVERSARIAL_COMPLETED = "ADVERSARIAL_COMPLETED"
DELTA_COMPUTED = "DELTA_COMPUTED"
# =============================================================================
# Benchmark Engine
# =============================================================================
class BenchmarkEngine:
"""
Main benchmarking engine for AegisLM.
Coordinates:
- Baseline evaluation (no attacks)
- Adversarial evaluation (full pipeline)
- Delta robustness computation
- Cross-model comparison
- Artifact generation
"""
def __init__(self):
self.logger = get_logger(__name__)
self._orchestrator = EvaluationOrchestrator()
self._active_benchmarks: Dict[str, asyncio.Task] = {}
def _log_event(
self,
event: BenchmarkEvent,
benchmark_id: str,
**kwargs: Any
) -> None:
"""Log benchmark event."""
log_data = {
"event": event.value,
"benchmark_id": benchmark_id,
"timestamp": datetime.utcnow().isoformat(),
}
log_data.update(kwargs)
if event in [BenchmarkEvent.BENCHMARK_STARTED, BenchmarkEvent.BENCHMARK_COMPLETED]:
self.logger.info("Benchmark event", **log_data)
elif event in [BenchmarkEvent.BENCHMARK_FAILED]:
self.logger.error("Benchmark event", **log_data)
else:
self.logger.debug("Benchmark event", **log_data)
async def start_benchmark(
self,
request: StartBenchmarkRequest,
) -> UUID:
"""
Start a new benchmark run.
Args:
request: Benchmark configuration
Returns:
Benchmark ID
"""
benchmark_id = uuid.uuid4()
benchmark_id_str = str(benchmark_id)
self.logger.info(
"Starting benchmark",
benchmark_id=benchmark_id_str,
models=request.models,
dataset=request.dataset_name,
)
# Create benchmark config
weights = request.weights or BenchmarkWeights()
config = BenchmarkConfig(
benchmark_id=benchmark_id,
models=request.models,
dataset_name=request.dataset_name,
dataset_version=request.dataset_version,
attack_enabled=request.attack_enabled,
mutation_depth=request.mutation_depth,
weights=weights,
max_concurrency=request.max_concurrency,
max_samples=request.max_samples,
enable_baseline=request.enable_baseline,
enable_adversarial=request.enable_adversarial,
attack_types=request.attack_types or ["jailbreak"],
)
# Validate config
config.validate_config()
# Start async execution
task = asyncio.create_task(
self._execute_benchmark(config)
)
self._active_benchmarks[benchmark_id_str] = task
return benchmark_id
async def _execute_benchmark(
self,
config: BenchmarkConfig,
) -> BenchmarkResult:
"""
Execute the benchmark asynchronously.
Args:
config: Benchmark configuration
Returns:
Complete benchmark result
"""
benchmark_id = config.benchmark_id
benchmark_id_str = str(benchmark_id)
start_time = datetime.utcnow()
# Initialize result
result = BenchmarkResult(
benchmark_id=benchmark_id,
dataset_name=config.dataset_name,
dataset_version=config.dataset_version,
models=config.models,
status=BenchmarkStatus.RUNNING,
results=[],
performance=BenchmarkPerformance(),
started_at=start_time,
config=config.model_dump(),
)
# Log start
self._log_event(
BenchmarkEvent.BENCHMARK_STARTED,
benchmark_id_str,
models=config.models,
dataset=config.dataset_name,
)
try:
# Evaluate each model
for model_name in config.models:
self._log_event(
BenchmarkEvent.MODEL_EVALUATION_STARTED,
benchmark_id_str,
model=model_name,
)
model_start_time = time.time()
# Evaluate model
model_result = await self._evaluate_model(
config=config,
model_name=model_name,
benchmark_id=benchmark_id_str,
)
model_time = time.time() - model_start_time
# Update performance tracking
result.performance.time_per_model_seconds[model_name] = model_time
result.performance.sample_counts[model_name] = (
model_result.adversarial.sample_count if model_result.adversarial else 0
)
result.performance.failure_rates[model_name] = (
model_result.adversarial.failure_rate if model_result.adversarial else 1.0
)
result.results.append(model_result)
self._log_event(
BenchmarkEvent.MODEL_EVALUATION_COMPLETED,
benchmark_id_str,
model=model_name,
time_seconds=model_time,
)
# Compute rankings (if multiple models)
if len(config.models) > 1:
result.rankings = rank_models(result.results)
# Generate vulnerability heatmap
result.vulnerability_heatmap = generate_vulnerability_heatmap(
result.results,
config.attack_types,
)
# Mark as completed
result.status = BenchmarkStatus.COMPLETED
result.completed_at = datetime.utcnow()
# Generate artifact
artifact_path = generate_benchmark_artifact(result)
self.logger.info(
"Benchmark artifact saved",
benchmark_id=benchmark_id_str,
path=artifact_path,
)
# Log completion
self._log_event(
BenchmarkEvent.BENCHMARK_COMPLETED,
benchmark_id_str,
models=config.models,
completed_at=result.completed_at.isoformat(),
)
except Exception as e:
result.status = BenchmarkStatus.FAILED
result.error = str(e)
result.completed_at = datetime.utcnow()
self.logger.error(
"Benchmark failed",
benchmark_id=benchmark_id_str,
error=str(e),
)
self._log_event(
BenchmarkEvent.BENCHMARK_FAILED,
benchmark_id_str,
error=str(e),
)
finally:
# Clean up active benchmark
self._active_benchmarks.pop(benchmark_id_str, None)
return result
async def _evaluate_model(
self,
config: BenchmarkConfig,
model_name: str,
benchmark_id: str,
) -> ModelBenchmarkResult:
"""
Evaluate a single model.
Args:
config: Benchmark configuration
model_name: Name of the model to evaluate
benchmark_id: Benchmark ID for logging
Returns:
Complete benchmark result for the model
"""
model_result = ModelBenchmarkResult(model_name=model_name)
# Create sampling config if max_samples is set
sampling_config = None
if config.max_samples:
sampling_config = {
"method": "random",
"sample_size": config.max_samples,
}
# Run baseline evaluation
if config.enable_baseline:
baseline_result = await self._run_evaluation(
model_name=model_name,
config=config,
mode=BenchmarkMode.BASELINE,
attack_enabled=False,
benchmark_id=benchmark_id,
sampling_config=sampling_config,
)
model_result.baseline = baseline_result
model_result.baseline_robustness = baseline_result.metrics.robustness
self._log_event(
BenchmarkEvent.BASELINE_COMPLETED,
benchmark_id,
model=model_name,
robustness=model_result.baseline_robustness,
)
# Run adversarial evaluation
if config.enable_adversarial:
adversarial_result = await self._run_evaluation(
model_name=model_name,
config=config,
mode=BenchmarkMode.ADVERSARIAL,
attack_enabled=config.attack_enabled,
benchmark_id=benchmark_id,
sampling_config=sampling_config,
)
model_result.adversarial = adversarial_result
model_result.adversarial_robustness = adversarial_result.metrics.robustness
self._log_event(
BenchmarkEvent.ADVERSARIAL_COMPLETED,
benchmark_id,
model=model_name,
robustness=model_result.adversarial_robustness,
)
# Compute deltas and derived metrics
if model_result.baseline and model_result.adversarial:
model_result.deltas = self._compute_deltas(
baseline=model_result.baseline,
adversarial=model_result.adversarial,
)
# Compute delta robustness
# ΔR = R_base - R_adv
model_result.delta_robustness = (
model_result.baseline_robustness - model_result.adversarial_robustness
)
# Compute Robustness Stability Index (RSI)
# RSI = R_adv / R_base
if model_result.baseline_robustness and model_result.baseline_robustness > 0:
model_result.robustness_stability_index = (
model_result.adversarial_robustness / model_result.baseline_robustness
)
else:
model_result.robustness_stability_index = 0.0
# Compute Vulnerability Index (VI)
# VI = delta_R / R_base
if model_result.baseline_robustness and model_result.baseline_robustness > 0:
model_result.vulnerability_index = (
model_result.delta_robustness / model_result.baseline_robustness
)
else:
model_result.vulnerability_index = 0.0
self._log_event(
BenchmarkEvent.DELTA_COMPUTED,
benchmark_id,
model=model_name,
delta_robustness=model_result.delta_robustness,
rsi=model_result.robustness_stability_index,
vi=model_result.vulnerability_index,
)
return model_result
async def _run_evaluation(
self,
model_name: str,
config: BenchmarkConfig,
mode: BenchmarkMode,
attack_enabled: bool,
benchmark_id: str,
sampling_config: Optional[Dict[str, Any]] = None,
) -> EvaluationResult:
"""
Run a single evaluation (baseline or adversarial).
Args:
model_name: Model to evaluate
config: Benchmark config
mode: Evaluation mode
attack_enabled: Whether to enable attacks
benchmark_id: Benchmark ID
sampling_config: Optional sampling config
Returns:
Evaluation result
"""
# Create evaluation input
eval_input = EvaluationInput(
model_name=model_name,
dataset_name=config.dataset_name,
dataset_version=config.dataset_version,
weights={
"hallucination": config.weights.hallucination,
"toxicity": config.weights.toxicity,
"bias": config.weights.bias,
"confidence": config.weights.confidence,
},
mutation_depth=config.mutation_depth if attack_enabled else 0,
attack_types=config.attack_types if attack_enabled else [],
max_concurrency=config.max_concurrency,
sampling_config=sampling_config,
)
# Run evaluation using orchestrator
output = await self._orchestrator.start_run(eval_input)
# Wait for completion
run_id = output.run_id
# Poll for completion (in production, this would be async callback)
max_wait = 300 # 5 minutes
waited = 0
poll_interval = 1
while waited < max_wait:
status = await self._orchestrator.get_run_status(run_id)
if status and status.status in [RunStatus.COMPLETED, RunStatus.FAILED]:
break
await asyncio.sleep(poll_interval)
waited += poll_interval
# Get final status
final_status = await self._orchestrator.get_run_status(run_id)
if final_status and final_status.status == RunStatus.COMPLETED:
# Extract metrics from output
metrics = ModelMetrics(
hallucination=final_status.metrics.get("hallucination", 0.5),
toxicity=final_status.metrics.get("toxicity", 0.5),
bias=final_status.metrics.get("bias", 0.5),
confidence=final_status.metrics.get("confidence", 0.5),
robustness=final_status.metrics.get("robustness", 0.5),
)
# Get standard deviations if available
if final_status.metrics:
metrics.std_hallucination = final_status.metrics.get("std_hallucination")
metrics.std_toxicity = final_status.metrics.get("std_toxicity")
metrics.std_bias = final_status.metrics.get("std_bias")
metrics.std_confidence = final_status.metrics.get("std_confidence")
return EvaluationResult(
model_name=model_name,
mode=mode,
metrics=metrics,
sample_count=final_status.metrics.get("total_samples", 0),
failure_rate=final_status.metrics.get("failed_samples", 0) / max(final_status.metrics.get("total_samples", 1), 1),
mean_latency_ms=final_status.performance.get("mean_latency_ms"),
total_time_seconds=final_status.performance.get("total_time_seconds"),
)
else:
# Return default result on failure
return EvaluationResult(
model_name=model_name,
mode=mode,
metrics=ModelMetrics(
hallucination=0.5,
toxicity=0.5,
bias=0.5,
confidence=0.5,
robustness=0.5,
),
sample_count=0,
failure_rate=1.0,
)
def _compute_deltas(
self,
baseline: EvaluationResult,
adversarial: EvaluationResult,
) -> MetricDeltas:
"""
Compute deltas between baseline and adversarial.
Args:
baseline: Baseline evaluation result
adversarial: Adversarial evaluation result
Returns:
MetricDeltas with computed differences
"""
return MetricDeltas(
hallucination_delta=adversarial.metrics.hallucination - baseline.metrics.hallucination,
toxicity_delta=adversarial.metrics.toxicity - baseline.metrics.toxicity,
bias_delta=adversarial.metrics.bias - baseline.metrics.bias,
confidence_delta=adversarial.metrics.confidence - baseline.metrics.confidence,
robustness_delta=baseline.metrics.robustness - adversarial.metrics.robustness,
)
async def get_benchmark_status(
self,
benchmark_id: str,
) -> Optional[BenchmarkResult]:
"""
Get status of a benchmark.
Args:
benchmark_id: Benchmark ID
Returns:
Benchmark result if found, None otherwise
"""
# Check if benchmark is active
if benchmark_id in self._active_benchmarks:
task = self._active_benchmarks[benchmark_id]
if not task.done():
# Benchmark is still running
# For now, return a partial result
return BenchmarkResult(
benchmark_id=UUID(benchmark_id),
dataset_name="",
dataset_version="",
models=[],
status=BenchmarkStatus.RUNNING,
results=[],
performance=BenchmarkPerformance(),
started_at=datetime.utcnow(),
)
else:
# Benchmark completed, get result
return await task
# Try to load from artifact
from backend.benchmarking.reporter import load_benchmark_artifact
artifact = load_benchmark_artifact(benchmark_id)
if artifact:
# Reconstruct BenchmarkResult from artifact
# For simplicity, just return None - in production, parse the artifact
pass
return None
async def cancel_benchmark(
self,
benchmark_id: str,
) -> bool:
"""
Cancel a running benchmark.
Args:
benchmark_id: Benchmark ID
Returns:
True if cancelled, False otherwise
"""
if benchmark_id in self._active_benchmarks:
task = self._active_benchmarks[benchmark_id]
task.cancel()
try:
await task
except asyncio.CancelledError:
pass
self.logger.info("Benchmark cancelled", benchmark_id=benchmark_id)
return True
return False
# =============================================================================
# Global Instance
# =============================================================================
_benchmark_engine: Optional[BenchmarkEngine] = None
def get_benchmark_engine() -> BenchmarkEngine:
"""
Get the global benchmark engine instance.
Returns:
BenchmarkEngine singleton
"""
global _benchmark_engine
if _benchmark_engine is None:
_benchmark_engine = BenchmarkEngine()
return _benchmark_engine
__all__ = [
"BenchmarkEngine",
"BenchmarkEvent",
"get_benchmark_engine",
]
|