Spaces:
Paused
Paused
File size: 17,733 Bytes
45ee481 | 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 | """
Human Evaluation Generator Module
Generate evaluation sets for blind testing by human evaluators.
Creates paired comparisons and rating tasks.
Example usage:
generator = HumanEvalGenerator()
eval_set = generator.generate_blind_test(
questions=test_questions,
responses_a=model_a_responses,
responses_b=model_b_responses,
)
"""
import json
import random
import hashlib
from dataclasses import dataclass, field
from datetime import datetime
from pathlib import Path
from typing import Optional
from loguru import logger
@dataclass
class EvaluationItem:
"""A single item for human evaluation."""
item_id: str
question: str
response_a: str
response_b: str
response_a_source: str # Hidden from evaluator
response_b_source: str # Hidden from evaluator
category: str
metadata: dict = field(default_factory=dict)
def to_evaluator_dict(self) -> dict:
"""Return dict for evaluator (without source info)."""
return {
"item_id": self.item_id,
"question": self.question,
"response_a": self.response_a,
"response_b": self.response_b,
"category": self.category,
}
def to_full_dict(self) -> dict:
"""Return full dict with sources."""
return {
"item_id": self.item_id,
"question": self.question,
"response_a": self.response_a,
"response_b": self.response_b,
"response_a_source": self.response_a_source,
"response_b_source": self.response_b_source,
"category": self.category,
"metadata": self.metadata,
}
@dataclass
class RatingItem:
"""A single item for absolute rating."""
item_id: str
question: str
response: str
source: str # Hidden
category: str
criteria: list = field(default_factory=list)
metadata: dict = field(default_factory=dict)
def to_evaluator_dict(self) -> dict:
"""Return dict for evaluator."""
return {
"item_id": self.item_id,
"question": self.question,
"response": self.response,
"category": self.category,
"criteria": self.criteria,
}
class HumanEvalGenerator:
"""
Generate evaluation sets for human evaluation.
Creates:
- Blind A/B comparisons
- Absolute rating tasks
- Multi-criteria evaluations
Example:
>>> generator = HumanEvalGenerator()
>>> eval_set = generator.generate_blind_test(questions, resp_a, resp_b)
>>> generator.save_evaluation_set(eval_set, "evaluation/")
"""
# Default evaluation criteria
DEFAULT_CRITERIA = [
{
"name": "voice_authenticity",
"description": "How well does the response capture the CEO's authentic voice and communication style?",
"scale": "1-5 (1=Not at all, 5=Perfectly authentic)",
},
{
"name": "helpfulness",
"description": "How helpful and substantive is the response in addressing the question?",
"scale": "1-5 (1=Not helpful, 5=Extremely helpful)",
},
{
"name": "clarity",
"description": "How clear and well-organized is the response?",
"scale": "1-5 (1=Confusing, 5=Crystal clear)",
},
{
"name": "professionalism",
"description": "How appropriate is the response for professional business communication?",
"scale": "1-5 (1=Inappropriate, 5=Highly professional)",
},
]
# Question categories for balanced evaluation
CATEGORIES = [
"opinion",
"strategic",
"factual",
"personal_philosophy",
"challenge",
]
def __init__(
self,
criteria: Optional[list] = None,
randomize: bool = True,
seed: Optional[int] = None,
):
"""
Initialize the generator.
Args:
criteria: Custom evaluation criteria
randomize: Whether to randomize response order
seed: Random seed for reproducibility
"""
self.criteria = criteria or self.DEFAULT_CRITERIA
self.randomize = randomize
if seed is not None:
random.seed(seed)
def generate_blind_test(
self,
questions: list[str],
responses_a: list[str],
responses_b: list[str],
source_a_name: str = "Model A",
source_b_name: str = "Model B",
categories: Optional[list[str]] = None,
) -> list[EvaluationItem]:
"""
Generate a blind A/B comparison test.
Args:
questions: List of questions
responses_a: Responses from source A
responses_b: Responses from source B
source_a_name: Name of source A (hidden from evaluator)
source_b_name: Name of source B (hidden from evaluator)
categories: Optional category for each question
Returns:
List of EvaluationItem objects
"""
if len(questions) != len(responses_a) or len(questions) != len(responses_b):
raise ValueError("Questions and responses must have same length")
categories = categories or ["general"] * len(questions)
items = []
for i, (q, ra, rb, cat) in enumerate(zip(questions, responses_a, responses_b, categories)):
# Generate unique ID
item_id = self._generate_id(q, ra, rb)
# Randomize order if enabled
if self.randomize and random.random() > 0.5:
ra, rb = rb, ra
source_a_name, source_b_name = source_b_name, source_a_name
items.append(EvaluationItem(
item_id=item_id,
question=q,
response_a=ra,
response_b=rb,
response_a_source=source_a_name,
response_b_source=source_b_name,
category=cat,
metadata={"index": i},
))
if self.randomize:
random.shuffle(items)
return items
def generate_rating_test(
self,
questions: list[str],
responses: list[str],
source_name: str = "Model",
categories: Optional[list[str]] = None,
criteria: Optional[list] = None,
) -> list[RatingItem]:
"""
Generate an absolute rating test.
Args:
questions: List of questions
responses: List of responses
source_name: Name of source (hidden from evaluator)
categories: Optional category for each question
criteria: Evaluation criteria to use
Returns:
List of RatingItem objects
"""
if len(questions) != len(responses):
raise ValueError("Questions and responses must have same length")
categories = categories or ["general"] * len(questions)
criteria = criteria or self.criteria
items = []
for i, (q, r, cat) in enumerate(zip(questions, responses, categories)):
item_id = self._generate_id(q, r)
items.append(RatingItem(
item_id=item_id,
question=q,
response=r,
source=source_name,
category=cat,
criteria=criteria,
metadata={"index": i},
))
if self.randomize:
random.shuffle(items)
return items
def generate_multi_model_test(
self,
questions: list[str],
model_responses: dict[str, list[str]],
categories: Optional[list[str]] = None,
) -> list[EvaluationItem]:
"""
Generate pairwise comparisons for multiple models.
Args:
questions: List of questions
model_responses: Dict of {model_name: responses}
categories: Optional categories
Returns:
List of pairwise EvaluationItem objects
"""
model_names = list(model_responses.keys())
if len(model_names) < 2:
raise ValueError("Need at least 2 models for comparison")
items = []
# Generate all pairs
for i in range(len(model_names)):
for j in range(i + 1, len(model_names)):
name_a = model_names[i]
name_b = model_names[j]
pair_items = self.generate_blind_test(
questions=questions,
responses_a=model_responses[name_a],
responses_b=model_responses[name_b],
source_a_name=name_a,
source_b_name=name_b,
categories=categories,
)
items.extend(pair_items)
if self.randomize:
random.shuffle(items)
return items
def save_evaluation_set(
self,
items: list,
output_dir: str | Path,
include_sources: bool = False,
) -> dict:
"""
Save evaluation set to files.
Args:
items: Evaluation items
output_dir: Output directory
include_sources: Include source info in evaluator file
Returns:
Dict with file paths
"""
output_dir = Path(output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
# Evaluator file (without source info)
evaluator_data = {
"metadata": {
"created": timestamp,
"num_items": len(items),
"criteria": self.criteria,
},
"items": [
item.to_evaluator_dict() for item in items
],
}
evaluator_path = output_dir / f"evaluation_set_{timestamp}.json"
with open(evaluator_path, "w", encoding="utf-8") as f:
json.dump(evaluator_data, f, indent=2, ensure_ascii=False)
# Answer key (with source info)
answer_key_data = {
"metadata": {
"created": timestamp,
"num_items": len(items),
},
"items": [
item.to_full_dict() for item in items
],
}
answer_key_path = output_dir / f"answer_key_{timestamp}.json"
with open(answer_key_path, "w", encoding="utf-8") as f:
json.dump(answer_key_data, f, indent=2, ensure_ascii=False)
# Instructions file
instructions = self._generate_instructions()
instructions_path = output_dir / "evaluation_instructions.md"
with open(instructions_path, "w", encoding="utf-8") as f:
f.write(instructions)
logger.info(f"Saved evaluation set to: {output_dir}")
return {
"evaluator_file": str(evaluator_path),
"answer_key": str(answer_key_path),
"instructions": str(instructions_path),
}
def load_results(
self,
results_path: str | Path,
answer_key_path: str | Path,
) -> dict:
"""
Load and analyze evaluation results.
Args:
results_path: Path to evaluator results
answer_key_path: Path to answer key
Returns:
Analysis results
"""
with open(results_path, "r") as f:
results = json.load(f)
with open(answer_key_path, "r") as f:
answer_key = json.load(f)
# Build lookup
key_lookup = {item["item_id"]: item for item in answer_key["items"]}
# Analyze
model_wins = {}
model_ratings = {}
for result in results.get("responses", []):
item_id = result.get("item_id")
if item_id not in key_lookup:
continue
key = key_lookup[item_id]
# For A/B comparisons
if "preference" in result:
pref = result["preference"]
if pref == "A":
winner = key.get("response_a_source", "A")
elif pref == "B":
winner = key.get("response_b_source", "B")
else:
winner = "tie"
model_wins[winner] = model_wins.get(winner, 0) + 1
# For ratings
if "ratings" in result:
source = key.get("source", "unknown")
if source not in model_ratings:
model_ratings[source] = {c["name"]: [] for c in self.criteria}
for criterion, rating in result["ratings"].items():
if criterion in model_ratings[source]:
model_ratings[source][criterion].append(rating)
# Calculate averages
avg_ratings = {}
for source, ratings in model_ratings.items():
avg_ratings[source] = {
criterion: sum(scores) / len(scores) if scores else 0
for criterion, scores in ratings.items()
}
return {
"comparison_wins": model_wins,
"average_ratings": avg_ratings,
"num_responses": len(results.get("responses", [])),
}
def _generate_id(self, *args) -> str:
"""Generate unique ID from content."""
content = "|".join(str(a)[:50] for a in args)
return hashlib.md5(content.encode()).hexdigest()[:12]
def _generate_instructions(self) -> str:
"""Generate evaluation instructions."""
criteria_text = "\n".join(
f"- **{c['name']}**: {c['description']} ({c['scale']})"
for c in self.criteria
)
return f"""# Human Evaluation Instructions
## Overview
You will be evaluating AI-generated responses to various questions. Your task is to assess the quality of these responses based on specific criteria.
## Evaluation Criteria
{criteria_text}
## Guidelines
1. **Read Carefully**: Read each question and response thoroughly before evaluating.
2. **Be Consistent**: Apply the same standards across all evaluations.
3. **Consider Context**: The responses are meant to represent a CEO's communication style for a technology company.
4. **A/B Comparisons**: When comparing two responses, select the one that better addresses the question while maintaining an authentic CEO voice.
5. **Rating Scale**: Use the full range of the rating scale. Reserve 5s for truly exceptional responses and 1s for clearly deficient ones.
## What to Look For
### Voice Authenticity
- Does it sound like a real CEO speaking?
- Is the tone confident but not arrogant?
- Does it reflect genuine experience and insight?
### Helpfulness
- Does it actually answer the question?
- Does it provide actionable insights?
- Is it substantive rather than generic?
### Clarity
- Is the response well-organized?
- Are ideas expressed clearly?
- Is it easy to follow the reasoning?
### Professionalism
- Is it appropriate for business communication?
- Does it maintain proper decorum?
- Is it culturally appropriate?
## Notes
- Take your time with each evaluation
- If unsure, re-read and consider both options carefully
- Your honest assessment is valuable
Thank you for your participation!
"""
def main():
"""CLI entry point for generating evaluation sets."""
import argparse
parser = argparse.ArgumentParser(
description="Generate human evaluation sets",
formatter_class=argparse.RawDescriptionHelpFormatter,
epilog="""
Examples:
python human_eval_generator.py --questions q.json --responses-a a.json --responses-b b.json
python human_eval_generator.py --questions q.json --responses r.json --rating-only
""",
)
parser.add_argument("--questions", required=True, help="Questions JSON file")
parser.add_argument("--responses-a", help="Responses A JSON file (for A/B test)")
parser.add_argument("--responses-b", help="Responses B JSON file (for A/B test)")
parser.add_argument("--responses", help="Responses JSON file (for rating test)")
parser.add_argument("--output", default="evaluation/", help="Output directory")
parser.add_argument("--rating-only", action="store_true", help="Generate rating test only")
parser.add_argument("--seed", type=int, help="Random seed")
args = parser.parse_args()
# Load questions
with open(args.questions, "r") as f:
questions_data = json.load(f)
questions = [q["question"] if isinstance(q, dict) else q for q in questions_data]
generator = HumanEvalGenerator(seed=args.seed)
if args.rating_only:
# Rating test
with open(args.responses, "r") as f:
responses_data = json.load(f)
responses = [r["response"] if isinstance(r, dict) else r for r in responses_data]
items = generator.generate_rating_test(questions, responses)
files = generator.save_evaluation_set(items, args.output)
else:
# A/B test
with open(args.responses_a, "r") as f:
resp_a_data = json.load(f)
with open(args.responses_b, "r") as f:
resp_b_data = json.load(f)
resp_a = [r["response"] if isinstance(r, dict) else r for r in resp_a_data]
resp_b = [r["response"] if isinstance(r, dict) else r for r in resp_b_data]
items = generator.generate_blind_test(questions, resp_a, resp_b)
files = generator.save_evaluation_set(items, args.output)
print(f"\nGenerated evaluation set:")
for name, path in files.items():
print(f" {name}: {path}")
return 0
if __name__ == "__main__":
exit(main())
|