File size: 55,543 Bytes
5374a2d |
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 |
# -----------------------------------------------------------------------------
# This file re-implements algorithms from the EvoPrompt project:
# Repo: GitHub - beeevita/EvoPrompt: Official implementation of the paper Connecting Large Language Models w
# Paper: "Connecting Large Language Models with Evolutionary Algorithms
# Yields Powerful Prompt Optimizers"
# Authors: Qingyan et al.
#
# Re-implementation integrated into EvoAgentX with permission from the authors.
# All mistakes or modifications are our own.
#
# Code of Conduct: This project follows the Microsoft Open Source Code of Conduct.
# https://opensource.microsoft.com/codeofconduct/
# -----------------------------------------------------------------------------
import asyncio
import random
import re
import os
import csv
import time
import itertools
from typing import Callable, Dict, List
from datetime import datetime
import numpy as np
from tqdm.asyncio import tqdm as aio_tqdm
import matplotlib.pyplot as plt
from evoagentx.agents import CustomizeAgent
from evoagentx.benchmark.bigbenchhard import BIGBenchHard
from evoagentx.core.logging import logger
from evoagentx.models import OpenAILLMConfig
from evoagentx.optimizers.engine.base import BaseOptimizer
from evoagentx.optimizers.engine.registry import ParamRegistry
class EvopromptOptimizer(BaseOptimizer):
"""
Base class for evolutionary prompt optimization algorithms.
This optimizer uses evolutionary algorithms to improve prompts in multi-agent workflows.
It supports both node-based and combination-based evolution strategies.
"""
def __init__(self,
registry: ParamRegistry,
program: Callable,
population_size: int,
iterations: int,
llm_config: OpenAILLMConfig,
concurrency_limit: int = 10,
combination_sample_size: int = None,
enable_logging: bool = True,
log_dir: str = None,
enable_early_stopping: bool = True,
early_stopping_patience: int = 3):
"""
Initialize the EvoPrompt optimizer.
Args:
registry: Parameter registry for tracking prompt nodes
program: The program/workflow to optimize
population_size: Size of the evolution population
iterations: Number of evolution iterations
llm_config: Configuration for the LLM used in evolution
concurrency_limit: Maximum concurrent API calls
combination_sample_size: Sample size for combination evaluation
enable_logging: Whether to enable detailed logging
log_dir: Directory for saving logs
enable_early_stopping: Whether to enable early stopping
early_stopping_patience: Number of generations to wait before stopping
"""
super().__init__(registry=registry, program=program)
# Core optimization parameters
self.population_size = population_size
self.iterations = iterations
self.llm_config = llm_config
self.semaphore = asyncio.Semaphore(concurrency_limit)
self.combination_sample_size = combination_sample_size
# Logging configuration
self.enable_logging = enable_logging
self.log_dir_base = log_dir
self.log_dir = None
# Early stopping mechanism
self.enable_early_stopping = enable_early_stopping
self.early_stopping_patience = early_stopping_patience
self._best_score_so_far = -float('inf')
self._generations_without_improvement = 0
# Evolution tracking data structures
self._eval_cache = {}
self.node_populations: Dict[str, List[str]] = {}
self.node_scores: Dict[str, List[float]] = {}
self.best_scores_per_gen: Dict[str, Dict[str, float]] = {}
self.avg_scores_per_gen: Dict[str, Dict[str, float]] = {}
self.best_combo_scores_per_gen: Dict[str, float] = {}
self.avg_combo_scores_per_gen: Dict[str, float] = {}
# Initialize paraphrase agent for prompt generation
self.paraphrase_agent = CustomizeAgent(
name="ParaphraseAgent",
description="An agent that paraphrases a given instruction.",
prompt="""Task: Generate a semantically equivalent but differently worded version of the user-provided instruction.
Now, please process the following instruction:
Input: {instruction}
Please provide the paraphrased version in the following format:
## paraphrased_instruction
[Your paraphrased version here]""",
llm_config=self.llm_config,
inputs=[
{"name": "instruction", "type": "string", "description": "The instruction to paraphrase."},
],
outputs=[
{"name": "paraphrased_instruction", "type": "string", "description": "The paraphrased instruction."}
],
parse_mode="title"
)
def _setup_logging_directory(self, benchmark: BIGBenchHard):
"""
Set up logging directory for evolution tracking.
Args:
benchmark: The benchmark instance containing task information
"""
if not self.enable_logging or self.log_dir:
return
task_name = benchmark.task if hasattr(benchmark, 'task') else 'unknown_task'
if self.log_dir_base is None:
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
algo_name = self.__class__.__name__.replace("Optimizer", "")
self.log_dir = f"node_evolution_logs_{algo_name}_{self.llm_config.model}_{task_name}_{timestamp}"
else:
self.log_dir = self.log_dir_base
os.makedirs(self.log_dir, exist_ok=True)
logger.info(f"Logging enabled. Log files will be saved to: {self.log_dir}")
def _log_generation_summary(self, generation: int, operation: str = "Evolution"):
"""
Log detailed summary of each generation's population and scores.
Args:
generation: The current generation number
operation: Type of operation (Evolution, Initial, etc.)
"""
if not self.enable_logging:
return
filename = f"generation_{generation:02d}_{operation.lower()}.csv"
filepath = os.path.join(self.log_dir, filename)
with open(filepath, 'w', newline='', encoding='utf-8') as f:
writer = csv.writer(f)
writer.writerow(['Node_Name', 'Individual_ID', 'Prompt_Text', 'Fitness_Score', 'Status', 'Rank_in_Node', 'Generation', 'Timestamp'])
timestamp = datetime.now().isoformat()
for node_name in self.node_populations.keys():
node_pop = self.node_populations.get(node_name, [])
node_scores = self.node_scores.get(node_name, [])
if not node_pop:
continue
sorted_indices = sorted(range(len(node_scores)), key=lambda i: node_scores[i], reverse=True)
for rank, idx in enumerate(sorted_indices, 1):
prompt = node_pop[idx]
score = node_scores[idx]
status = "Best" if rank == 1 else "Survivor" if rank <= self.population_size else "Eliminated"
writer.writerow([
node_name, f"{node_name}_{idx}", prompt[:200] + "..." if len(prompt) > 200 else prompt,
f"{score:.6f}", status, rank, generation, timestamp
])
def _log_detailed_evaluation(self, generation: int, combinations: List[Dict[str, str]],
combination_scores: List[float]):
if not self.enable_logging:
return
filename = f"combo_evaluation_gen_{generation:02d}.csv"
filepath = os.path.join(self.log_dir, filename)
with open(filepath, 'w', newline='', encoding='utf-8') as f:
writer = csv.writer(f)
node_names = list(combinations[0].keys()) if combinations else []
header = ['Combination_ID', 'Average_Score']
for node_name in node_names:
header.append(f'{node_name}_Prompt_Preview')
header.extend(['Generation', 'Timestamp'])
writer.writerow(header)
timestamp = datetime.now().isoformat()
for combo_id, (combination, avg_score) in enumerate(zip(combinations, combination_scores)):
try:
row = [f"combo_{combo_id}", f"{avg_score:.6f}"]
for node_name in node_names:
prompt = combination[node_name]
row.append(prompt[:50] + "..." if len(prompt) > 50 else prompt)
row.extend([generation, timestamp])
writer.writerow(row)
except Exception as e:
logger.error(f"Error logging evaluation for combination {combo_id}: {e}")
def _create_single_metric_plot(self, metric_name: str, generations: List[int],
best_scores: List[float], avg_scores: List[float],
algorithm_name: str, plot_dir: str):
fig, ax = plt.subplots(figsize=(12, 7))
ax.plot(generations, best_scores, marker='o', linestyle='-', linewidth=2, markersize=8, label='Best Score')
ax.plot(generations, avg_scores, marker='x', linestyle='--', linewidth=2, markersize=8, label='Average Score')
title = f"Performance for '{metric_name}' ({algorithm_name})"
ax.set_title(title, fontsize=16, weight='bold')
ax.set_xlabel('Generation', fontsize=12)
ax.set_ylabel('Fitness Score', fontsize=12)
ax.set_xticks(generations)
ax.set_xticklabels([f"Gen {g}" for g in generations], rotation=45, ha="right")
ax.legend(loc='best', fontsize=10)
ax.grid(True, which='both', linestyle='--', linewidth=0.5)
plt.tight_layout()
safe_metric_name = re.sub(r'[^a-zA-Z0-9_-]', '_', metric_name)
filename = f"performance_plot_{safe_metric_name}.png"
filepath = os.path.join(plot_dir, filename)
try:
plt.savefig(filepath, dpi=200, bbox_inches='tight')
except Exception as e:
logger.error(f"Failed to save individual plot for {metric_name}: {e}")
finally:
plt.close(fig)
def _plot_and_save_performance_graph(self, algorithm_name: str):
if not self.enable_logging or plt is None:
if plt is None:
logger.warning("Matplotlib not found, skipping plot generation.")
return
if not self.best_scores_per_gen and not self.best_combo_scores_per_gen:
logger.warning("No performance data to plot.")
return
plt.style.use('seaborn-v0_8-whitegrid')
all_gen_keys = set(self.best_scores_per_gen.keys()) | set(self.best_combo_scores_per_gen.keys())
generations = sorted([int(re.search(r'\d+', gen).group()) for gen in all_gen_keys if re.search(r'\d+', gen)])
fig_combined, ax_combined = plt.subplots(figsize=(16, 9))
if self.best_combo_scores_per_gen:
combo_best = [self.best_combo_scores_per_gen.get(f"Gen_{g}") for g in generations]
combo_avg = [self.avg_combo_scores_per_gen.get(f"Gen_{g}") for g in generations]
ax_combined.plot(generations, combo_best, marker='*', linestyle='-', linewidth=2.5, markersize=10, label='Best Combination Score (Overall)')
ax_combined.plot(generations, combo_avg, marker='D', linestyle='--', linewidth=2.5, markersize=8, label='Average Combination Score (Overall)')
all_node_metrics = set()
for gen_data in self.best_scores_per_gen.values():
all_node_metrics.update(gen_data.keys())
for metric in sorted(list(all_node_metrics)):
best_scores = [self.best_scores_per_gen.get(f"Gen_{g}", {}).get(metric) for g in generations]
avg_scores = [self.avg_scores_per_gen.get(f"Gen_{g}", {}).get(metric) for g in generations]
ax_combined.plot(generations, best_scores, marker='o', linestyle='-', alpha=0.7, label=f'Best Score ({metric})')
ax_combined.plot(generations, avg_scores, marker='x', linestyle='--', alpha=0.7, label=f'Average Score ({metric})')
ax_combined.set_title(f'Overall Performance Evolution ({algorithm_name})', fontsize=18, weight='bold')
ax_combined.set_xlabel('Generation', fontsize=14)
ax_combined.set_ylabel('Fitness Score', fontsize=14)
ax_combined.set_xticks(generations)
ax_combined.set_xticklabels([f"Gen {g}" for g in generations], rotation=45, ha="right")
handles, labels = ax_combined.get_legend_handles_labels()
combo_indices = [i for i, label in enumerate(labels) if 'Combination' in label]
node_indices = [i for i, label in enumerate(labels) if 'Combination' not in label]
ax_combined.legend([handles[i] for i in combo_indices + node_indices],
[labels[i] for i in combo_indices + node_indices],
loc='best', fontsize=10)
ax_combined.grid(True, which='both', linestyle='--', linewidth=0.5)
plt.tight_layout()
combined_filepath = os.path.join(self.log_dir, "performance_summary_OVERALL.png")
try:
plt.savefig(combined_filepath, dpi=300, bbox_inches='tight')
logger.info(f"Overall performance plot saved to: {combined_filepath}")
except Exception as e:
logger.error(f"Failed to save overall performance plot: {e}")
finally:
plt.close(fig_combined)
individual_plot_dir = os.path.join(self.log_dir, 'individual_plots')
os.makedirs(individual_plot_dir, exist_ok=True)
for metric in sorted(list(all_node_metrics)):
best_scores = [self.best_scores_per_gen.get(f"Gen_{g}", {}).get(metric) for g in generations]
avg_scores = [self.avg_scores_per_gen.get(f"Gen_{g}", {}).get(metric) for g in generations]
self._create_single_metric_plot(metric, generations, best_scores, avg_scores, algorithm_name, individual_plot_dir)
if self.best_combo_scores_per_gen:
combo_best = [self.best_combo_scores_per_gen.get(f"Gen_{g}") for g in generations]
combo_avg = [self.avg_combo_scores_per_gen.get(f"Gen_{g}") for g in generations]
self._create_single_metric_plot("Combination", generations, combo_best, combo_avg, algorithm_name, individual_plot_dir)
logger.info(f"Individual performance plots saved to: {individual_plot_dir}")
def _log_optimization_summary(self, algorithm_name: str, best_config: Dict[str, str],
test_accuracy: float = None):
if not self.enable_logging:
return
filename = f"optimization_summary_{algorithm_name.lower()}.csv"
filepath = os.path.join(self.log_dir, filename)
with open(filepath, 'w', newline='', encoding='utf-8') as f:
writer = csv.writer(f)
writer.writerow(['Metric', 'Value', 'Timestamp'])
timestamp = datetime.now().isoformat()
writer.writerow(['Algorithm', algorithm_name, timestamp])
writer.writerow(['Population_Size', self.population_size, timestamp])
writer.writerow(['Iterations', self.iterations, timestamp])
writer.writerow(['Combination_Sample_Size', self.combination_sample_size, timestamp])
writer.writerow(['Early_Stopping_Enabled', self.enable_early_stopping, timestamp])
if self.enable_early_stopping:
writer.writerow(['Early_Stopping_Patience', self.early_stopping_patience, timestamp])
if test_accuracy is not None:
writer.writerow(['Final_Test_Accuracy', f"{test_accuracy:.6f}", timestamp])
for node_name, prompt in best_config.items():
writer.writerow([f'Best_{node_name}', prompt, timestamp])
for gen_name in self.best_scores_per_gen.keys():
for metric_name, best_score in self.best_scores_per_gen[gen_name].items():
writer.writerow([f'{gen_name}_{metric_name}_Best', f"{best_score:.6f}", timestamp])
if gen_name in self.avg_scores_per_gen:
for metric_name, avg_score in self.avg_scores_per_gen[gen_name].items():
writer.writerow([f'{gen_name}_{metric_name}_Avg', f"{avg_score:.6f}", timestamp])
self._plot_and_save_performance_graph(algorithm_name)
async def _log_evaluation_details(self, benchmark: BIGBenchHard, dataset: List[Dict],
predictions: List[str], scores: List[float], eval_mode: str,
accuracy: float, correct_count: int, total_count: int):
if not self.enable_logging:
return
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
filename = f"evaluation_testset_{eval_mode}_{timestamp}.csv"
filepath = os.path.join(self.log_dir, filename)
logger.info(f"Logging detailed evaluation results to {filepath}")
with open(filepath, 'w', newline='', encoding='utf-8') as f:
writer = csv.writer(f)
writer.writerow(['Metric', 'Value'])
writer.writerow(['Overall_Accuracy', f"{accuracy:.6f}"])
writer.writerow(['Correct_Count', correct_count])
writer.writerow(['Total_Count', total_count])
writer.writerow([])
# Write detailed data header
writer.writerow(['example_id', 'input_text', 'prediction', 'ground_truth', 'score'])
for i, example in enumerate(dataset):
example_id = benchmark._get_id(example)
input_text = example.get("input", "")
label = benchmark.get_label(example)
writer.writerow([
example_id,
input_text[:200] + "..." if len(input_text) > 200 else input_text,
predictions[i],
label,
scores[i]
])
def _log_generation(self, generation: int, combos_with_scores: List[tuple]):
"""
Log generation data for combination-based evolution.
"""
if not self.enable_logging:
return
filename = f"combo_generation_{generation:02d}_log.csv"
filepath = os.path.join(self.log_dir, filename)
with open(filepath, 'w', newline='', encoding='utf-8') as f:
writer = csv.writer(f)
header = ['Combination_ID', 'Combination_Score', 'Node_Name', 'Prompt_Text', 'Generation', 'Timestamp']
writer.writerow(header)
timestamp = datetime.now().isoformat()
sorted_combos = sorted(combos_with_scores, key=lambda x: x[1], reverse=True)
for combo_rank, (combination, avg_score) in enumerate(sorted_combos):
combo_id = f"combo_rank_{combo_rank + 1}"
for node_name, prompt_text in combination.items():
writer.writerow([
combo_id,
f"{avg_score:.6f}",
node_name,
prompt_text[:200] + "..." if len(prompt_text) > 200 else prompt_text,
generation,
timestamp
])
async def _evaluate_combination_list(self, combinations: List[Dict], benchmark: BIGBenchHard, dev_set: list) -> List[float]:
if not combinations:
return []
eval_dev_set = dev_set[:50] if len(dev_set) > 50 else dev_set
all_scores = []
pbar = aio_tqdm(total=len(combinations), desc="Evaluating batch", leave=False)
for combo in combinations:
tasks = [self._evaluate_combination_on_example(combo, benchmark, ex) for ex in eval_dev_set]
example_scores = await asyncio.gather(*tasks)
avg_score = sum(example_scores) / len(example_scores) if example_scores else 0.0
all_scores.append(avg_score)
pbar.update(1)
pbar.close()
return all_scores
def _generate_combinations(self, node_populations: Dict[str, List[str]]) -> List[Dict[str, str]]:
node_names = list(node_populations.keys())
node_prompts = [node_populations[node] for node in node_names]
total_possible = np.prod([len(p) for p in node_prompts if p]) if all(p for p in node_prompts) else 0
if total_possible == 0:
logger.warning("Cannot generate combinations, one or more node populations are empty.")
return []
if self.combination_sample_size is None:
target_size = min(self.population_size, int(total_possible), 200)
else:
target_size = min(self.combination_sample_size, int(total_possible))
logger.info(f"Total possible combinations: {total_possible}, sampling: {target_size}")
if target_size >= total_possible:
all_combinations = []
for combination in itertools.product(*node_prompts):
combo_dict = {node_names[i]: combination[i] for i in range(len(node_names))}
all_combinations.append(combo_dict)
return all_combinations
sampled_combinations = []
sampled_keys = set()
max_attempts = target_size * 5
attempts = 0
while len(sampled_combinations) < target_size and attempts < max_attempts:
combination = {name: random.choice(prompts) for name, prompts in node_populations.items()}
combo_key = tuple(sorted(combination.items()))
if combo_key not in sampled_keys:
sampled_combinations.append(combination)
sampled_keys.add(combo_key)
attempts += 1
logger.info(f"Generated {len(sampled_combinations)} unique combinations")
return sampled_combinations
async def _evaluate_combination_on_example(self, combination: Dict[str, str],
benchmark: BIGBenchHard, example: Dict) -> float:
combo_key = tuple(sorted(combination.items()))
example_key = str(hash(str(example)))
cache_key = hash((combo_key, example_key))
if not hasattr(self, '_eval_cache'):
self._eval_cache = {}
if cache_key in self._eval_cache:
return self._eval_cache[cache_key]
async with self.semaphore:
try:
original_config = self.get_current_cfg()
self.apply_cfg(combination)
inputs = {k: v for k, v in example.items() if k in benchmark.get_input_keys()}
prediction, _ = await asyncio.to_thread(self.program, **inputs)
label = benchmark.get_label(example)
score_dict = benchmark.evaluate(prediction, label)
score = score_dict.get("em", 0.0)
self.apply_cfg(original_config)
self._eval_cache[cache_key] = score
if len(self._eval_cache) > 5000:
keys_to_del = list(self._eval_cache.keys())[:1000]
for key in keys_to_del:
del self._eval_cache[key]
return score
except Exception as e:
logger.error(f"Error evaluating combination: {e}")
return 0.0
async def _evaluate_combinations_and_update_node_scores(self, combinations: List[Dict[str, str]],
benchmark: BIGBenchHard, dev_set: list) -> List[float]:
eval_dev_set = dev_set[:50] if len(dev_set) > 50 else dev_set
combination_scores = []
print(f"Evaluating {len(combinations)} combinations on {len(eval_dev_set)} examples...")
combo_pbar = aio_tqdm(total=len(combinations), desc="Evaluating Combinations")
for combination in combinations:
tasks = [self._evaluate_combination_on_example(combination, benchmark, ex) for ex in eval_dev_set]
example_scores = await asyncio.gather(*tasks)
avg_score = sum(example_scores) / len(example_scores) if example_scores else 0.0
combination_scores.append(avg_score)
combo_pbar.update(1)
combo_pbar.close()
for node_name in self.node_populations.keys():
self.node_scores[node_name] = [0.0] * len(self.node_populations[node_name])
for prompt_idx, prompt in enumerate(self.node_populations[node_name]):
participating_scores = [
combo_score for combo_idx, combo_score in enumerate(combination_scores)
if combinations[combo_idx].get(node_name) == prompt
]
if participating_scores:
self.node_scores[node_name][prompt_idx] = sum(participating_scores) / len(participating_scores)
else:
self.node_scores[node_name][prompt_idx] = 0.0
return combination_scores
async def _perform_paraphrase(self, prompt: str) -> str:
async with self.semaphore:
output = await asyncio.to_thread(
self.paraphrase_agent,
inputs={"instruction": prompt}
)
return output.content.paraphrased_instruction.strip()
async def _perform_evolution(self, agent: Callable, inputs: Dict[str, str]) -> str:
async with self.semaphore:
output = await asyncio.to_thread(agent, inputs=inputs)
if hasattr(output.content, 'evolved_prompt'):
return output.content.evolved_prompt.strip()
return str(output.content).strip()
async def _initialize_node_populations(self, initial_config: Dict[str, any]):
for node_name, initial_value in initial_config.items():
node_population = []
if isinstance(initial_value, list):
provided_size = len(initial_value)
if self.population_size < provided_size:
logger.info(f"Node '{node_name}': Provided population ({provided_size}) is larger than target size ({self.population_size}). Randomly sampling.")
node_population = random.sample(initial_value, self.population_size)
elif self.population_size == provided_size:
logger.info(f"Node '{node_name}': Provided population size ({provided_size}) matches target size. Using directly.")
node_population = list(initial_value)
else:
logger.info(f"Node '{node_name}': Target population size ({self.population_size}) is larger than provided ({provided_size}). Expanding.")
node_population = list(initial_value)
num_to_generate = self.population_size - provided_size
source_prompts_for_generation = random.choices(initial_value, k=num_to_generate)
paraphrase_tasks = [self._perform_paraphrase(prompt) for prompt in source_prompts_for_generation]
new_prompts = await aio_tqdm.gather(
*paraphrase_tasks, desc=f"Expanding population for {node_name}"
)
node_population.extend(new_prompts)
elif isinstance(initial_value, str):
logger.info(f"Node '{node_name}': Generating population from a single initial prompt.")
node_population = [initial_value]
if self.population_size > 1:
num_to_generate = self.population_size - 1
paraphrase_tasks = [self._perform_paraphrase(initial_value) for _ in range(num_to_generate)]
new_prompts = await aio_tqdm.gather(
*paraphrase_tasks, desc=f"Generating initial population for {node_name}"
)
node_population.extend(new_prompts)
else:
raise TypeError(f"Unsupported type for tracked parameter '{node_name}': {type(initial_value)}. Must be str or list.")
self.node_populations[node_name] = node_population
self.node_scores[node_name] = [0.0] * self.population_size
async def evaluate(self, benchmark: BIGBenchHard, eval_mode: str = "test") -> Dict[str, float]:
"""
Evaluates the optimized program on a specified dataset.
Args:
benchmark (BIGBenchHard): The benchmark instance containing the data.
eval_mode (str): The evaluation mode, either "test" or "dev".
Returns:
Dict[str, float]: A dictionary containing evaluation metrics.
"""
logger.info(f"--- Evaluating optimized program on '{eval_mode}' set ---")
dataset = benchmark.get_test_data() if eval_mode == "test" else benchmark.get_dev_data()
if not dataset:
logger.warning(f"No data found for '{eval_mode}' set. Returning empty results.")
return {}
async def evaluate_example(example: Dict) -> tuple[float, str]:
prediction, _ = await asyncio.to_thread(self.program, input=example["input"])
score_dict = benchmark.evaluate(prediction, benchmark.get_label(example))
score = score_dict.get("em", 0.0)
return score, prediction
tasks = [evaluate_example(ex) for ex in dataset]
results = await aio_tqdm.gather(*tasks, desc=f"Evaluating on {eval_mode.capitalize()} Set")
scores, predictions = zip(*results) if results else ([], [])
correct_count = sum(scores)
total_count = len(dataset)
accuracy = correct_count / total_count if total_count > 0 else 0.0
logger.info(f"{eval_mode.capitalize()} Set Accuracy: {accuracy:.4f} ({int(correct_count)}/{total_count})")
if self.enable_logging:
await self._log_evaluation_details(
benchmark, dataset, predictions, scores, eval_mode,
accuracy, int(correct_count), total_count
)
return {"accuracy": accuracy}
class GAOptimizer(EvopromptOptimizer):
"""
Genetic Algorithm optimizer for prompt evolution.
This optimizer uses genetic algorithm operations (crossover, mutation, selection)
to evolve prompts. It supports both node-based and combination-based evolution.
"""
def __init__(self, *args, full_evaluation: bool = False, **kwargs):
"""
Initialize the GA optimizer.
Args:
full_evaluation: Whether to use full node-based evaluation or combination-based
*args: Arguments passed to parent class
**kwargs: Keyword arguments passed to parent class
"""
super().__init__(*args, **kwargs)
self.full_evaluation = full_evaluation
# Log initialization mode
mode_str = "full_evaluation" if self.full_evaluation else "combination-based"
logger.info(f"GAOptimizer initialized with '{mode_str}' mode.")
# Initialize genetic algorithm agent for prompt evolution
self.ga_agent = CustomizeAgent(
name="ga_agent",
description="An agent that evolves a new prompt from two parent prompts.",
prompt="""Please follow the instructions step-by-step to generate a better prompt.
1. Crossover the following prompts to generate a new prompt:
Prompt 1: {parent1}
Prompt 2: {parent2}
2. Mutate the prompt generated in Step 1 and generate a final evolved prompt. Strictly preserve the original XML tags structure.
Now process the given prompts and provide your output in the following format:
## evolved_prompt
[Your evolved prompt here]""",
llm_config=self.llm_config,
inputs=[
{"name": "parent1", "type": "string", "description": "The first parent prompt."},
{"name": "parent2", "type": "string", "description": "The second parent prompt."}
],
outputs=[
{"name": "evolved_prompt", "type": "string", "description": "The evolved prompt with XML tags preserved."}
],
parse_mode="title"
)
async def _perform_node_evolution(self, node_name: str, node_population: List[str],
node_scores: List[float] = None,
evolution_agent: Callable = None) -> List[str]:
probabilities = None
if node_scores:
total_fitness = sum(node_scores)
if total_fitness > 0:
probabilities = [s / total_fitness for s in node_scores]
agent = evolution_agent or self.ga_agent
num_children_to_create = len(node_population)
evolution_tasks = []
for _ in range(num_children_to_create):
parents = random.choices(node_population, weights=probabilities, k=2) if probabilities else random.choices(node_population, k=2)
task = self._perform_evolution(agent=agent, inputs={"parent1": parents[0], "parent2": parents[1]})
evolution_tasks.append(task)
new_children = await aio_tqdm.gather(*evolution_tasks, desc=f"Evolving {node_name}")
return new_children
async def optimize(self, benchmark: BIGBenchHard) -> tuple[Dict[str, str], dict, dict]:
self._setup_logging_directory(benchmark)
initial_config = self.get_current_cfg()
if not initial_config:
raise ValueError("Registry is empty.")
await self._initialize_node_populations(initial_config)
dev_set = benchmark.get_dev_data()
if not dev_set:
raise ValueError("Benchmark has no development set.")
self._best_score_so_far = -float('inf')
self._generations_without_improvement = 0
if self.full_evaluation:
logger.info("--- Starting Node-Based Evolution with Makeup Evaluation (full_evaluation=True) ---")
print("--- Step 1: Initial evaluation of node combinations... ---")
combinations = self._generate_combinations(self.node_populations)
combination_scores = await self._evaluate_combinations_and_update_node_scores(combinations, benchmark, dev_set)
self._log_generation_summary(0, "Initial")
self._log_detailed_evaluation(0, combinations, combination_scores)
self.best_scores_per_gen["Gen_0"] = {name: max(scores) if scores else 0 for name, scores in self.node_scores.items()}
self.avg_scores_per_gen["Gen_0"] = {name: np.mean(scores) if scores else 0 for name, scores in self.node_scores.items()}
if combination_scores:
initial_best_combo_score = max(combination_scores)
self._best_score_so_far = initial_best_combo_score
self.best_combo_scores_per_gen["Gen_0"] = initial_best_combo_score
self.avg_combo_scores_per_gen["Gen_0"] = np.mean(combination_scores)
logger.info(f"Early stopping baseline set to initial best combination score: {self._best_score_so_far:.4f}")
for t in range(self.iterations):
generation_start_time = time.time()
print(f"\n--- Generation {t + 1}/{self.iterations} ---")
children_populations = {}
for node_name in self.node_populations.keys():
children = await self._perform_node_evolution(
node_name, self.node_populations[node_name], self.node_scores[node_name], self.ga_agent
)
children_populations[node_name] = children
current_populations = {
name: self.node_populations[name] + children_populations[name]
for name in self.node_populations.keys()
}
self.node_populations = current_populations
print(f"Performing main evaluation for {len(list(current_populations.values())[0])} individuals in each node...")
combinations = self._generate_combinations(self.node_populations)
combination_scores = await self._evaluate_combinations_and_update_node_scores(combinations, benchmark, dev_set)
prompts_needing_makeup = []
for node_name, scores in self.node_scores.items():
for idx, score in enumerate(scores):
if score == 0.0:
prompt_to_check = self.node_populations[node_name][idx]
is_in_combos = any(c.get(node_name) == prompt_to_check for c in combinations)
if not is_in_combos:
prompts_needing_makeup.append((node_name, idx, prompt_to_check))
if prompts_needing_makeup:
print(f"--- Performing makeup evaluation for {len(prompts_needing_makeup)} unsampled individuals... ---")
makeup_combinations = []
for node_name, idx, prompt in prompts_needing_makeup:
makeup_combo = {name: random.choice(pop) for name, pop in self.node_populations.items()}
makeup_combo[node_name] = prompt
makeup_combinations.append(makeup_combo)
makeup_scores = await self._evaluate_combination_list(makeup_combinations, benchmark, dev_set)
for i, (node_name, idx, prompt) in enumerate(prompts_needing_makeup):
self.node_scores[node_name][idx] = makeup_scores[i]
logger.info(f"Updated score for '{prompt[:30]}...' to {makeup_scores[i]:.4f} after makeup eval.")
print("--- Selecting survivors for the next generation... ---")
survivor_populations = {}
survivor_scores = {}
for node_name in self.node_populations.keys():
population = self.node_populations[node_name]
scores = self.node_scores[node_name]
sorted_pairs = sorted(zip(scores, population), key=lambda x: x[0], reverse=True)
selected_pairs = sorted_pairs[:self.population_size]
if selected_pairs:
selected_scores, selected_population = zip(*selected_pairs)
survivor_scores[node_name] = list(selected_scores)
survivor_populations[node_name] = list(selected_population)
else:
survivor_scores[node_name], survivor_populations[node_name] = [], []
print(f"Node {node_name}: Selected top {len(survivor_populations[node_name])} from {len(population)} individuals")
self.node_populations = survivor_populations
self.node_scores = survivor_scores
generation_time = time.time() - generation_start_time
print(f"Generation {t + 1} completed in {generation_time:.2f}s")
self._log_generation_summary(t + 1, "Evolution")
if combination_scores:
self._log_detailed_evaluation(t + 1, combinations, combination_scores)
gen_name = f"Gen_{t + 1}"
self.best_scores_per_gen[gen_name] = {name: max(scores) if scores else 0 for name, scores in self.node_scores.items()}
self.avg_scores_per_gen[gen_name] = {name: np.mean(scores) if scores else 0 for name, scores in self.node_scores.items()}
best_combo_score_this_gen = max(combination_scores) if combination_scores else -float('inf')
self.best_combo_scores_per_gen[gen_name] = best_combo_score_this_gen
self.avg_combo_scores_per_gen[gen_name] = np.mean(combination_scores) if combination_scores else 0.0
if self.enable_early_stopping:
if best_combo_score_this_gen > self._best_score_so_far + 1e-6:
self._best_score_so_far = best_combo_score_this_gen
self._generations_without_improvement = 0
logger.info(f"Early stopping: New best combination score found: {self._best_score_so_far:.4f}.")
else:
self._generations_without_improvement += 1
logger.info(f"Early stopping: No improvement for {self._generations_without_improvement} generation(s).")
if self._generations_without_improvement >= self.early_stopping_patience:
logger.warning(f"\n--- EARLY STOPPING TRIGGERED at generation {t + 1} ---")
break
else:
logger.info("--- Starting Combo-Based Evolution (full_evaluation=False) ---")
print("--- Step 1: Creating and evaluating initial combination population... ---")
initial_combinations = self._generate_combinations(self.node_populations)
initial_scores = await self._evaluate_combination_list(initial_combinations, benchmark, dev_set)
combo_population_with_scores = sorted(zip(initial_combinations, initial_scores), key=lambda x: x[1], reverse=True)
combo_population_with_scores = combo_population_with_scores[:self.population_size]
gen_0_scores = [score for _, score in combo_population_with_scores]
if gen_0_scores:
best_gen_score = max(gen_0_scores)
avg_gen_score = np.mean(gen_0_scores)
self.best_combo_scores_per_gen["Gen_0"] = best_gen_score
self.avg_combo_scores_per_gen["Gen_0"] = avg_gen_score
self._best_score_so_far = best_gen_score
print(f"Generation 0 complete. Best score: {best_gen_score:.4f}, Avg score: {avg_gen_score:.4f}")
logger.info(f"Early stopping baseline set to: {self._best_score_so_far:.4f}")
self._log_generation(0, combo_population_with_scores)
for t in range(self.iterations):
print(f"\n--- Generation {t + 1}/{self.iterations} (Combo Evolution) ---")
parent_prompts_for_node = {name: [] for name in initial_config.keys()}
for combo, _ in combo_population_with_scores:
for node_name, prompt in combo.items():
parent_prompts_for_node[node_name].append(prompt)
children_populations = {}
for node_name, prompts in parent_prompts_for_node.items():
children_populations[node_name] = await self._perform_node_evolution(node_name, prompts)
print("Evaluating new child combinations...")
child_combinations = self._generate_combinations(children_populations)
child_scores = await self._evaluate_combination_list(child_combinations, benchmark, dev_set)
child_combos_with_scores = list(zip(child_combinations, child_scores))
print("Selecting best combinations from parents and children...")
combined_population = combo_population_with_scores + child_combos_with_scores
sorted_combos = sorted(combined_population, key=lambda x: x[1], reverse=True)
combo_population_with_scores = sorted_combos[:self.population_size]
self._log_generation(t + 1, combo_population_with_scores)
current_scores = [score for _, score in combo_population_with_scores]
best_gen_score = max(current_scores) if current_scores else 0
avg_gen_score = np.mean(current_scores) if current_scores else 0
gen_name = f"Gen_{t + 1}"
self.best_combo_scores_per_gen[gen_name] = best_gen_score
self.avg_combo_scores_per_gen[gen_name] = avg_gen_score
print(f"Generation {t + 1} complete. Best score: {best_gen_score:.4f}, Avg score: {avg_gen_score:.4f}")
if self.enable_early_stopping:
if best_gen_score > self._best_score_so_far + 1e-6:
self._best_score_so_far = best_gen_score
self._generations_without_improvement = 0
logger.info(f"Early stopping: New best combination score found: {self._best_score_so_far:.4f}. Patience counter reset.")
else:
self._generations_without_improvement += 1
logger.info(f"Early stopping: No improvement for {self._generations_without_improvement} generation(s). Patience: {self.early_stopping_patience}.")
if self._generations_without_improvement >= self.early_stopping_patience:
logger.warning(f"\n--- EARLY STOPPING TRIGGERED at generation {t + 1} ---")
break
print("\n--- Evolution complete ---")
if self.full_evaluation:
best_config = {
name: self.node_populations[name][np.argmax(self.node_scores[name])]
for name in self.node_populations.keys() if self.node_populations.get(name) and self.node_scores.get(name)
}
else:
best_config, _ = max(combo_population_with_scores, key=lambda x: x[1]) if combo_population_with_scores else ({}, 0)
self._log_optimization_summary("GA", best_config)
self.apply_cfg(best_config)
logger.info("Optimization finished! The best configuration has been applied to the program.")
return best_config, self.best_combo_scores_per_gen, self.avg_scores_per_gen
class DEOptimizer(EvopromptOptimizer):
"""
Differential Evolution optimizer for prompt evolution.
This optimizer uses differential evolution strategy for prompt optimization,
including mutation, crossover, and selection operations based on DE principles.
"""
def __init__(self, *args, **kwargs):
"""
Initialize the DE optimizer.
Args:
*args: Arguments passed to parent class
**kwargs: Keyword arguments passed to parent class
"""
super().__init__(*args, **kwargs)
# Initialize differential evolution agent for prompt mutation
self.de_agent = CustomizeAgent(
name="DE_Agent",
description="Generates a new trial prompt using the Differential Evolution strategy.",
prompt="""Please follow the instructions step-by-step to generate a better prompt using Differential Evolution strategy.
1. Identify the different parts between these two donor prompts:
Donor Prompt 1: {donor1}
Donor Prompt 2: {donor2}
2. Randomly mutate the different parts identified above.
3. Combine the mutated parts with the best prompt, selectively replacing its content:
Best Prompt: {best_prompt}
4. Crossover the result from Step 3 with the current prompt to generate the final evolved prompt. Strictly preserve the original XML tags structure:
Current Prompt: {current_prompt}
Please provide the final evolved prompt in the following format:
## evolved_prompt
[Your evolved prompt here]""",
llm_config=self.llm_config,
inputs=[
{"name": "current_prompt", "type": "string", "description": "The base prompt to be mutated, p_i."},
{"name": "donor1", "type": "string", "description": "The first donor prompt, p_r1."},
{"name": "donor2", "type": "string", "description": "The second donor prompt, p_r2."},
{"name": "best_prompt", "type": "string", "description": "The best prompt found so far in the population, p_best."},
],
outputs=[
{"name": "evolved_prompt", "type": "string", "description": "The evolved prompt with XML tags preserved."}
],
parse_mode="title"
)
async def _evolve_and_select_one(
self,
target_combo_with_score: tuple,
full_pop_with_scores: List[tuple],
benchmark: BIGBenchHard,
dev_set: list
) -> tuple:
"""
Evolve a single combination using differential evolution and select the better one.
Args:
target_combo_with_score: The target combination and its score
full_pop_with_scores: The full population with scores
benchmark: The benchmark for evaluation
dev_set: Development set for evaluation
Returns:
Tuple of the better combination (target or trial) and its score
"""
target_combo, target_score = target_combo_with_score
best_combo, _ = max(full_pop_with_scores, key=lambda x: x[1])
# Select donor combinations (avoid using target as donor)
donor_pool = [c for c in full_pop_with_scores if c[0] != target_combo]
if len(donor_pool) < 2:
donors = random.choices(full_pop_with_scores, k=2)
else:
donors = random.sample(donor_pool, 2)
donor1_combo, _ = donors[0]
donor2_combo, _ = donors[1]
# Evolve each prompt node using DE mutation and crossover
evolution_tasks = []
node_names = list(target_combo.keys())
for node_name in node_names:
task = self._perform_evolution(
agent=self.de_agent,
inputs={
"current_prompt": target_combo[node_name],
"donor1": donor1_combo[node_name],
"donor2": donor2_combo[node_name],
"best_prompt": best_combo[node_name]
}
)
evolution_tasks.append(task)
# Evaluate trial combination and perform selection
evolved_components = await asyncio.gather(*evolution_tasks)
trial_combo = {name: comp for name, comp in zip(node_names, evolved_components)}
trial_scores = await self._evaluate_combination_list([trial_combo], benchmark, dev_set)
trial_score = trial_scores[0]
# Selection: return the better combination
return (trial_combo, trial_score) if trial_score > target_score else (target_combo, target_score)
async def optimize(self, benchmark: BIGBenchHard) -> tuple[Dict[str, str], dict, dict]:
self._setup_logging_directory(benchmark)
initial_config = self.get_current_cfg()
if not initial_config:
raise ValueError("Registry is empty.")
logger.info("Optimizing with DEOptimizer (Pipelined Combination Evolution).")
await self._initialize_node_populations(initial_config)
dev_set = benchmark.get_dev_data()
if not dev_set:
raise ValueError("Benchmark has no development set.")
self._best_score_so_far = -float('inf')
self._generations_without_improvement = 0
print("--- Step 1: Creating and evaluating initial combination population... ---")
initial_combinations = self._generate_combinations(self.node_populations)
initial_scores = await self._evaluate_combination_list(initial_combinations, benchmark, dev_set)
combo_pop_with_scores = list(zip(initial_combinations, initial_scores))
self._log_generation(0, combo_pop_with_scores)
initial_best = max(initial_scores) if initial_scores else 0
initial_avg = np.mean(initial_scores) if initial_scores else 0
self.best_combo_scores_per_gen["Gen_0"] = initial_best
self.avg_combo_scores_per_gen["Gen_0"] = initial_avg
print(f"Initial population - Best score: {initial_best:.4f}, Avg score: {initial_avg:.4f}")
if initial_scores:
self._best_score_so_far = initial_best
for t in range(self.iterations):
print(f"\n--- Generation {t + 1}/{self.iterations} ---")
evolution_pipeline_tasks = [
self._evolve_and_select_one(combo_with_score, combo_pop_with_scores, benchmark, dev_set)
for combo_with_score in combo_pop_with_scores
]
pbar = aio_tqdm(total=len(evolution_pipeline_tasks), desc=f"Pipelined Evolution Gen {t+1}")
next_gen_pop_with_scores = []
for future in asyncio.as_completed(evolution_pipeline_tasks):
result = await future
next_gen_pop_with_scores.append(result)
pbar.update(1)
pbar.close()
combo_pop_with_scores = next_gen_pop_with_scores
self._log_generation(t + 1, combo_pop_with_scores)
current_scores = [score for _, score in combo_pop_with_scores]
best_gen_score = max(current_scores) if current_scores else 0
avg_gen_score = np.mean(current_scores) if current_scores else 0
gen_name = f"Gen_{t + 1}"
self.best_combo_scores_per_gen[gen_name] = best_gen_score
self.avg_combo_scores_per_gen[gen_name] = avg_gen_score
print(f"Generation {t + 1} complete. Best score: {best_gen_score:.4f}, Avg score: {avg_gen_score:.4f}")
if self.enable_early_stopping:
if best_gen_score > self._best_score_so_far + 1e-6:
self._best_score_so_far = best_gen_score
self._generations_without_improvement = 0
logger.info(f"Early stopping: New best score found: {self._best_score_so_far:.4f}. Patience counter reset.")
else:
self._generations_without_improvement += 1
logger.info(f"Early stopping: No improvement for {self._generations_without_improvement} generation(s). Patience: {self.early_stopping_patience}.")
if self._generations_without_improvement >= self.early_stopping_patience:
logger.warning(f"\n--- EARLY STOPPING TRIGGERED at generation {t + 1} ---")
logger.warning(f"No improvement in best score for {self.early_stopping_patience} consecutive generations.")
break
print("\n--- Combination-Level Evolution complete ---")
best_combination, best_score = max(combo_pop_with_scores, key=lambda x: x[1]) if combo_pop_with_scores else ({}, 0)
logger.info(f"Optimization finished! Best combination found with score {best_score:.4f}.")
self._log_optimization_summary("DE", best_combination)
self.apply_cfg(best_combination)
return best_combination, self.best_combo_scores_per_gen, self.avg_combo_scores_per_gen |