File size: 15,973 Bytes
b684ab3 |
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 |
#!/usr/bin/env python3
"""
Goodhart Gap Benchmark Evaluation Script
Evaluate any model on the Goodhart Gap benchmark to detect the gap
between understanding and execution in multi-step reasoning.
Usage:
# Using OpenAI API
python evaluate.py --provider openai --model gpt-4o
# Using Anthropic API
python evaluate.py --provider anthropic --model claude-3-5-haiku-latest
# Using local Ollama
python evaluate.py --provider ollama --model llama3.1:8b
# Using HuggingFace transformers
python evaluate.py --provider huggingface --model meta-llama/Llama-3.1-8B-Instruct
# Custom API endpoint
python evaluate.py --provider custom --model mymodel --api-url http://localhost:8000/v1
Environment Variables:
OPENAI_API_KEY - Required for OpenAI provider
ANTHROPIC_API_KEY - Required for Anthropic provider
HF_TOKEN - Optional for gated HuggingFace models
"""
import argparse
import json
import os
import re
import sys
from dataclasses import dataclass
from datetime import datetime
from pathlib import Path
from typing import Optional, Callable
import time
# Optional imports
try:
import requests
HAS_REQUESTS = True
except ImportError:
HAS_REQUESTS = False
@dataclass
class TestResult:
id: str
domain: str
problem: str
expected: str
response: str
extracted_answer: str
passed: bool
latency_ms: float
def extract_answer(response: str, expected: str) -> str:
"""Extract the answer from model response."""
response = response.strip()
# Try to find numbers in the response
numbers = re.findall(r'-?[\d,]+\.?\d*', response)
# For yes/no questions
if expected.lower() in ['yes', 'no']:
resp_lower = response.lower()
if 'yes' in resp_lower and 'no' not in resp_lower.split()[:3]:
return 'yes'
if 'no' in resp_lower and 'yes' not in resp_lower.split()[:3]:
return 'no'
if 'cannot determine' in resp_lower or 'cannot be determined' in resp_lower:
return 'cannot determine'
# For time answers
time_match = re.search(r'(\d{1,2}:\d{2})\s*(AM|PM|am|pm)?', response)
if time_match:
time_str = time_match.group(1)
period = time_match.group(2) or ''
return f"{time_str} {period}".strip()
# For ordering questions (comma-separated names)
if ',' in expected and any(c.isalpha() for c in expected):
# Try to extract comma-separated list
parts = [p.strip() for p in response.split(',') if p.strip()]
if len(parts) >= 3:
return ', '.join(parts[:5])
# Return first number found
if numbers:
return numbers[0].replace(',', '')
# Return first line or truncated response
first_line = response.split('\n')[0]
return first_line[:50] if len(first_line) > 50 else first_line
def validate_answer(response: str, expected: str, domain: str) -> bool:
"""Validate if the response matches the expected answer."""
response = response.lower().strip()
expected = expected.lower().strip()
# Direct match
if expected in response:
return True
# Numeric comparison
expected_nums = re.findall(r'-?[\d,]+\.?\d*', expected)
response_nums = re.findall(r'-?[\d,]+\.?\d*', response)
if expected_nums and response_nums:
try:
exp_val = float(expected_nums[0].replace(',', ''))
for resp_num in response_nums:
resp_val = float(resp_num.replace(',', ''))
# Allow small floating point tolerance
if abs(exp_val - resp_val) < 0.01:
return True
# Check if it's within 0.5% (for rounding)
if exp_val != 0 and abs(exp_val - resp_val) / abs(exp_val) < 0.005:
return True
except ValueError:
pass
# Time validation
if domain == 'time':
# Normalize time formats
def normalize_time(t):
t = t.lower().replace(' ', '')
t = re.sub(r'(\d{1,2}):(\d{2})(am|pm)?', r'\1:\2\3', t)
return t
if normalize_time(expected) in normalize_time(response):
return True
# Yes/no validation
if expected in ['yes', 'no', 'cannot determine']:
if expected == 'yes' and 'yes' in response and 'no' not in response.split()[:5]:
return True
if expected == 'no' and 'no' in response and 'yes' not in response.split()[:5]:
return True
if expected == 'cannot determine' and ('cannot' in response or 'unable' in response):
return True
# Ordering validation (check sequence)
if ',' in expected and domain == 'logic':
expected_items = [x.strip().lower() for x in expected.split(',')]
response_lower = response.lower()
# Check if items appear in correct order
positions = []
for item in expected_items:
pos = response_lower.find(item)
if pos == -1:
return False
positions.append(pos)
return positions == sorted(positions)
return False
class ModelProvider:
"""Base class for model providers."""
def generate(self, prompt: str) -> tuple[str, float]:
"""Generate response. Returns (response, latency_ms)."""
raise NotImplementedError
class OpenAIProvider(ModelProvider):
def __init__(self, model: str, api_key: Optional[str] = None):
self.model = model
self.api_key = api_key or os.environ.get('OPENAI_API_KEY')
if not self.api_key:
raise ValueError("OPENAI_API_KEY not set")
def generate(self, prompt: str) -> tuple[str, float]:
start = time.time()
headers = {
"Authorization": f"Bearer {self.api_key}",
"Content-Type": "application/json"
}
payload = {
"model": self.model,
"messages": [{"role": "user", "content": prompt}],
"temperature": 0.1,
"max_tokens": 200
}
response = requests.post(
"https://api.openai.com/v1/chat/completions",
headers=headers, json=payload, timeout=60
)
latency = (time.time() - start) * 1000
if response.status_code == 200:
return response.json()["choices"][0]["message"]["content"].strip(), latency
else:
return f"ERROR: {response.status_code}", latency
class AnthropicProvider(ModelProvider):
def __init__(self, model: str, api_key: Optional[str] = None):
self.model = model
self.api_key = api_key or os.environ.get('ANTHROPIC_API_KEY')
if not self.api_key:
raise ValueError("ANTHROPIC_API_KEY not set")
def generate(self, prompt: str) -> tuple[str, float]:
start = time.time()
headers = {
"x-api-key": self.api_key,
"anthropic-version": "2023-06-01",
"Content-Type": "application/json"
}
payload = {
"model": self.model,
"max_tokens": 200,
"messages": [{"role": "user", "content": prompt}]
}
response = requests.post(
"https://api.anthropic.com/v1/messages",
headers=headers, json=payload, timeout=60
)
latency = (time.time() - start) * 1000
if response.status_code == 200:
return response.json()["content"][0]["text"].strip(), latency
else:
return f"ERROR: {response.status_code}", latency
class OllamaProvider(ModelProvider):
def __init__(self, model: str, host: str = "http://localhost:11434"):
self.model = model
self.host = host
def generate(self, prompt: str) -> tuple[str, float]:
start = time.time()
payload = {
"model": self.model,
"prompt": prompt,
"stream": False,
"options": {"temperature": 0.1}
}
response = requests.post(
f"{self.host}/api/generate",
json=payload, timeout=120
)
latency = (time.time() - start) * 1000
if response.status_code == 200:
return response.json().get("response", "").strip(), latency
else:
return f"ERROR: {response.status_code}", latency
class CustomProvider(ModelProvider):
def __init__(self, model: str, api_url: str):
self.model = model
self.api_url = api_url
def generate(self, prompt: str) -> tuple[str, float]:
start = time.time()
# Assume OpenAI-compatible API
payload = {
"model": self.model,
"messages": [{"role": "user", "content": prompt}],
"temperature": 0.1,
"max_tokens": 200
}
response = requests.post(
f"{self.api_url}/chat/completions",
json=payload, timeout=120
)
latency = (time.time() - start) * 1000
if response.status_code == 200:
return response.json()["choices"][0]["message"]["content"].strip(), latency
else:
return f"ERROR: {response.status_code}", latency
def load_dataset(path: str = "data/test.jsonl") -> list[dict]:
"""Load the benchmark dataset."""
problems = []
with open(path) as f:
for line in f:
problems.append(json.loads(line))
return problems
def evaluate_model(
provider: ModelProvider,
problems: list[dict],
verbose: bool = False
) -> tuple[list[TestResult], dict]:
"""Evaluate a model on the benchmark."""
results = []
domain_stats = {}
for i, problem in enumerate(problems):
if verbose:
print(f"[{i+1}/{len(problems)}] {problem['id']}...", end=" ", flush=True)
response, latency = provider.generate(problem['problem'])
extracted = extract_answer(response, problem['correct_answer'])
passed = validate_answer(response, problem['correct_answer'], problem['domain'])
result = TestResult(
id=problem['id'],
domain=problem['domain'],
problem=problem['problem'],
expected=problem['correct_answer'],
response=response[:200],
extracted_answer=extracted,
passed=passed,
latency_ms=latency
)
results.append(result)
# Track domain stats
domain = problem['domain']
if domain not in domain_stats:
domain_stats[domain] = {'pass': 0, 'fail': 0}
domain_stats[domain]['pass' if passed else 'fail'] += 1
if verbose:
status = "PASS" if passed else "FAIL"
print(f"{status} (got: {extracted[:20]})")
# Calculate summary
total_pass = sum(r.passed for r in results)
total = len(results)
summary = {
'total': total,
'passed': total_pass,
'failed': total - total_pass,
'pass_rate': total_pass / total if total > 0 else 0,
'by_domain': {
d: {
'passed': s['pass'],
'total': s['pass'] + s['fail'],
'pass_rate': s['pass'] / (s['pass'] + s['fail'])
}
for d, s in domain_stats.items()
},
'avg_latency_ms': sum(r.latency_ms for r in results) / len(results) if results else 0
}
return results, summary
def save_results(
results: list[TestResult],
summary: dict,
model_name: str,
output_dir: str = "results"
):
"""Save evaluation results."""
os.makedirs(output_dir, exist_ok=True)
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
safe_model = re.sub(r'[^\w\-]', '_', model_name)
# Save detailed results
results_file = f"{output_dir}/{safe_model}_{timestamp}_results.jsonl"
with open(results_file, 'w') as f:
for r in results:
f.write(json.dumps({
'id': r.id,
'domain': r.domain,
'expected': r.expected,
'response': r.response,
'extracted': r.extracted_answer,
'passed': r.passed,
'latency_ms': r.latency_ms
}) + '\n')
# Save summary
summary_file = f"{output_dir}/{safe_model}_{timestamp}_summary.json"
summary['model'] = model_name
summary['timestamp'] = timestamp
with open(summary_file, 'w') as f:
json.dump(summary, f, indent=2)
return results_file, summary_file
def print_summary(summary: dict, model_name: str):
"""Print evaluation summary."""
print("\n" + "=" * 60)
print(f"GOODHART GAP BENCHMARK RESULTS")
print(f"Model: {model_name}")
print("=" * 60)
print(f"\nOverall: {summary['passed']}/{summary['total']} ({summary['pass_rate']*100:.1f}%)")
print(f"Average latency: {summary['avg_latency_ms']:.0f}ms")
print("\nBy Domain:")
print("-" * 40)
for domain, stats in sorted(summary['by_domain'].items()):
bar = "█" * int(stats['pass_rate'] * 10) + "░" * (10 - int(stats['pass_rate'] * 10))
print(f" {domain:<15} {stats['passed']:>2}/{stats['total']:<2} {bar} {stats['pass_rate']*100:>5.1f}%")
print("\n" + "=" * 60)
# Interpret results
pass_rate = summary['pass_rate']
if pass_rate >= 0.9:
print("Assessment: LOW GOODHART GAP - Model executes well")
elif pass_rate >= 0.7:
print("Assessment: MODERATE GOODHART GAP - Some execution issues")
elif pass_rate >= 0.5:
print("Assessment: SIGNIFICANT GOODHART GAP - Frequent execution failures")
else:
print("Assessment: SEVERE GOODHART GAP - Major execution problems")
def main():
parser = argparse.ArgumentParser(
description="Evaluate a model on the Goodhart Gap Benchmark",
formatter_class=argparse.RawDescriptionHelpFormatter,
epilog=__doc__
)
parser.add_argument('--provider', required=True,
choices=['openai', 'anthropic', 'ollama', 'custom'],
help='Model provider')
parser.add_argument('--model', required=True,
help='Model name/identifier')
parser.add_argument('--api-url', default=None,
help='API URL for custom provider')
parser.add_argument('--data', default='data/test.jsonl',
help='Path to test data')
parser.add_argument('--output', default='results',
help='Output directory')
parser.add_argument('--verbose', '-v', action='store_true',
help='Show progress')
parser.add_argument('--limit', type=int, default=None,
help='Limit number of problems (for testing)')
args = parser.parse_args()
if not HAS_REQUESTS:
print("ERROR: requests library required. Install with: pip install requests")
sys.exit(1)
# Create provider
if args.provider == 'openai':
provider = OpenAIProvider(args.model)
elif args.provider == 'anthropic':
provider = AnthropicProvider(args.model)
elif args.provider == 'ollama':
provider = OllamaProvider(args.model)
elif args.provider == 'custom':
if not args.api_url:
print("ERROR: --api-url required for custom provider")
sys.exit(1)
provider = CustomProvider(args.model, args.api_url)
# Load dataset
print(f"Loading dataset from {args.data}...")
problems = load_dataset(args.data)
if args.limit:
problems = problems[:args.limit]
print(f"Loaded {len(problems)} problems")
# Evaluate
print(f"\nEvaluating {args.model}...")
results, summary = evaluate_model(provider, problems, verbose=args.verbose)
# Save and print results
results_file, summary_file = save_results(results, summary, args.model, args.output)
print_summary(summary, args.model)
print(f"\nResults saved to:")
print(f" {results_file}")
print(f" {summary_file}")
if __name__ == "__main__":
main()
|