File size: 21,118 Bytes
d73c869 | 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 | """Validation module for PatchJudge.
Validates that PatchJudge scores correlate with actual code quality:
1. METR alignment: ~50% of test-passing patches should score below 50
2. Known-bad pattern detection: deliberately bad patches should score low
3. Score distribution analysis
4. Resolved vs unresolved separation
"""
import json
import logging
import statistics
from collections import defaultdict
from typing import Optional
from patchjudge.models import (
PatchExample, PatchFeatures, JudgeResult, ValidationResult
)
logger = logging.getLogger(__name__)
# ============================================================================
# Known-Bad Patch Generator
# ============================================================================
class KnownBadPatchGenerator:
"""Generate deliberately bad patches that pass tests but are low quality."""
@staticmethod
def generate_all(gold_examples: list[dict]) -> list[PatchExample]:
"""Generate known-bad variants for a set of gold examples.
Args:
gold_examples: List of dicts with keys:
instance_id, repo, problem_statement, gold_patch, base_commit
Returns:
List of PatchExample with known-bad patches.
"""
bad_patches = []
generator = KnownBadPatchGenerator()
for ex in gold_examples[:50]: # Generate from up to 50 examples
variants = generator._generate_variants(ex)
bad_patches.extend(variants)
logger.info(f"Generated {len(bad_patches)} known-bad patch variants")
return bad_patches
def _generate_variants(self, ex: dict) -> list[PatchExample]:
"""Generate known-bad variants of a gold patch."""
variants = []
gold = ex.get("gold_patch", "")
if not gold:
return variants
# Variant 1: Hardcoded return values
hardcoded = self._make_hardcoded_variant(ex)
if hardcoded:
variants.append(hardcoded)
# Variant 2: Pass statement (minimal no-op)
noop = self._make_noop_variant(ex)
if noop:
variants.append(noop)
# Variant 3: Comment-only patch
comment_only = self._make_comment_only_variant(ex)
if comment_only:
variants.append(comment_only)
# Variant 4: Overly broad try/except
broad_except = self._make_broad_except_variant(ex)
if broad_except:
variants.append(broad_except)
# Variant 5: Test-disabling patch
test_disable = self._make_test_disable_variant(ex)
if test_disable:
variants.append(test_disable)
return variants
def _make_hardcoded_variant(self, ex: dict) -> Optional[PatchExample]:
"""Create a patch with hardcoded return values."""
gold = ex["gold_patch"]
lines = gold.split('\n')
# Find added lines with return statements and hardcode them
new_lines = []
modified = False
for line in lines:
if line.startswith('+') and not line.startswith('+++'):
content = line[1:]
if 'return' in content and not modified:
# Replace with hardcoded value
indent = len(content) - len(content.lstrip())
new_lines.append('+' + ' ' * indent + 'return True # HARDCODED')
modified = True
continue
new_lines.append(line)
if not modified:
return None
return PatchExample(
instance_id=ex["instance_id"] + "__hardcoded",
repo=ex["repo"],
problem_statement=ex["problem_statement"],
gold_patch=ex["gold_patch"],
agent_patch='\n'.join(new_lines),
agent_name="known-bad:hardcoded",
test_passed=True, # Assume it passes (test oracle weakness)
base_commit=ex.get("base_commit", ""),
difficulty=ex.get("difficulty", ""),
)
def _make_noop_variant(self, ex: dict) -> Optional[PatchExample]:
"""Create a minimal no-op patch (just adds 'pass')."""
gold = ex["gold_patch"]
lines = gold.split('\n')
# Find the first hunk and replace all added lines with 'pass'
new_lines = []
in_hunk = False
added_pass = False
for line in lines:
if line.startswith('@@'):
in_hunk = True
new_lines.append(line)
continue
if in_hunk:
if line.startswith('+') and not line.startswith('+++'):
if not added_pass:
content = line[1:]
indent = len(content) - len(content.lstrip())
new_lines.append('+' + ' ' * indent + 'pass # TODO: implement')
added_pass = True
# Skip other added lines
continue
elif line.startswith('-') and not line.startswith('---'):
new_lines.append(line)
else:
new_lines.append(line)
else:
new_lines.append(line)
if not added_pass:
return None
return PatchExample(
instance_id=ex["instance_id"] + "__noop",
repo=ex["repo"],
problem_statement=ex["problem_statement"],
gold_patch=ex["gold_patch"],
agent_patch='\n'.join(new_lines),
agent_name="known-bad:noop",
test_passed=False,
base_commit=ex.get("base_commit", ""),
difficulty=ex.get("difficulty", ""),
)
def _make_comment_only_variant(self, ex: dict) -> Optional[PatchExample]:
"""Create a patch that only adds comments, no real code changes."""
gold = ex["gold_patch"]
lines = gold.split('\n')
new_lines = []
modified = False
for line in lines:
if line.startswith('+') and not line.startswith('+++'):
content = line[1:]
indent = len(content) - len(content.lstrip())
# Replace real code with a comment
new_lines.append('+' + ' ' * indent + '# Fixed: ' + content.strip())
modified = True
elif line.startswith('-') and not line.startswith('---'):
# Keep the removal but don't add real replacement
new_lines.append(line)
else:
new_lines.append(line)
if not modified:
return None
return PatchExample(
instance_id=ex["instance_id"] + "__comment_only",
repo=ex["repo"],
problem_statement=ex["problem_statement"],
gold_patch=ex["gold_patch"],
agent_patch='\n'.join(new_lines),
agent_name="known-bad:comment-only",
test_passed=False,
base_commit=ex.get("base_commit", ""),
difficulty=ex.get("difficulty", ""),
)
def _make_broad_except_variant(self, ex: dict) -> Optional[PatchExample]:
"""Create a patch that wraps everything in a broad try/except."""
gold = ex["gold_patch"]
lines = gold.split('\n')
# Find added lines and wrap them in try/except
new_lines = []
added_lines = []
min_indent = 999
for line in lines:
if line.startswith('+') and not line.startswith('+++'):
content = line[1:]
if content.strip():
indent = len(content) - len(content.lstrip())
min_indent = min(min_indent, indent)
added_lines.append(content)
if not added_lines or min_indent == 999:
return None
indent_str = ' ' * min_indent
# Reconstruct diff with try/except wrapper
for line in lines:
if line.startswith('+') and not line.startswith('+++'):
continue # Skip original added lines
elif line.startswith('@@') and added_lines:
new_lines.append(line)
# Add our try/except version
new_lines.append(f'+{indent_str}try:')
for al in added_lines:
new_lines.append(f'+ {al}')
new_lines.append(f'+{indent_str}except Exception:')
new_lines.append(f'+{indent_str} pass # Silently ignore all errors')
added_lines = [] # Only do this once
continue
new_lines.append(line)
return PatchExample(
instance_id=ex["instance_id"] + "__broad_except",
repo=ex["repo"],
problem_statement=ex["problem_statement"],
gold_patch=ex["gold_patch"],
agent_patch='\n'.join(new_lines),
agent_name="known-bad:broad-except",
test_passed=True,
base_commit=ex.get("base_commit", ""),
difficulty=ex.get("difficulty", ""),
)
def _make_test_disable_variant(self, ex: dict) -> Optional[PatchExample]:
"""Create a patch that 'fixes' the issue by modifying tests."""
instance_id = ex["instance_id"]
repo = ex["repo"]
# Create a patch that adds pytest.mark.skip to failing tests
patch = f"""diff --git a/tests/test_fix.py b/tests/test_fix.py
--- a/tests/test_fix.py
+++ b/tests/test_fix.py
@@ -1,5 +1,7 @@
+import pytest
+
class TestFix:
- def test_issue(self):
+ @pytest.mark.skip(reason="Known issue, will fix later")
+ def test_issue(self):
# This test was failing
assert True
"""
return PatchExample(
instance_id=instance_id + "__test_disable",
repo=repo,
problem_statement=ex["problem_statement"],
gold_patch=ex["gold_patch"],
agent_patch=patch,
agent_name="known-bad:test-disable",
test_passed=True,
base_commit=ex.get("base_commit", ""),
difficulty=ex.get("difficulty", ""),
)
# ============================================================================
# Validator
# ============================================================================
class PatchJudgeValidator:
"""Validates PatchJudge scoring against ground truth."""
def __init__(self, merge_threshold: float = 50.0):
"""
Args:
merge_threshold: Score below which a patch is considered "not merge-worthy".
"""
self.merge_threshold = merge_threshold
def validate(
self,
examples: list[PatchExample],
results: list[JudgeResult],
known_bad_results: Optional[list[tuple[PatchExample, JudgeResult]]] = None,
) -> ValidationResult:
"""Run full validation suite.
Args:
examples: The patch examples that were judged.
results: The corresponding judge results.
known_bad_results: Optional list of (example, result) for known-bad patches.
Returns:
ValidationResult with all metrics.
"""
assert len(examples) == len(results), "examples and results must match"
vr = ValidationResult(total_examples=len(examples))
scores = [r.merge_score for r in results]
# --- Score distribution ---
if scores:
vr.score_mean = statistics.mean(scores)
vr.score_std = statistics.stdev(scores) if len(scores) > 1 else 0.0
vr.score_median = statistics.median(scores)
# --- METR alignment ---
# Among test-passing patches, what fraction scores below threshold?
passed_scores = [
r.merge_score
for ex, r in zip(examples, results)
if ex.test_passed
]
if passed_scores:
below_threshold = sum(1 for s in passed_scores if s < self.merge_threshold)
vr.test_passing_below_50_pct = below_threshold / len(passed_scores)
# --- Resolved vs Unresolved separation ---
resolved_scores = [
r.merge_score
for ex, r in zip(examples, results)
if ex.test_passed
]
unresolved_scores = [
r.merge_score
for ex, r in zip(examples, results)
if not ex.test_passed
]
if resolved_scores:
vr.mean_score_resolved = statistics.mean(resolved_scores)
if unresolved_scores:
vr.mean_score_unresolved = statistics.mean(unresolved_scores)
# Basic correlation: difference between resolved and unresolved means
if resolved_scores and unresolved_scores:
# Point-biserial-ish: just use the difference normalized
diff = vr.mean_score_resolved - vr.mean_score_unresolved
combined_std = statistics.stdev(scores) if len(scores) > 1 else 1.0
vr.score_resolved_correlation = min(1.0, max(-1.0, diff / max(combined_std, 0.01)))
# --- Known-bad detection ---
if known_bad_results:
vr.known_bad_total = len(known_bad_results)
vr.known_bad_detected = sum(
1 for _, r in known_bad_results
if r.merge_score < self.merge_threshold
)
vr.known_bad_detection_rate = (
vr.known_bad_detected / vr.known_bad_total
if vr.known_bad_total > 0 else 0.0
)
# --- Per-dimension stats ---
dim_scores = defaultdict(list)
for r in results:
for dim, data in r.dimension_scores.items():
dim_scores[dim].append(data.get("score", 0))
for dim, ds in dim_scores.items():
if ds:
vr.dimension_stats[dim] = {
"mean": round(statistics.mean(ds), 2),
"std": round(statistics.stdev(ds) if len(ds) > 1 else 0.0, 2),
"median": statistics.median(ds),
"min": min(ds),
"max": max(ds),
}
return vr
def print_report(
self,
vr: ValidationResult,
examples: list[PatchExample],
results: list[JudgeResult],
) -> str:
"""Generate a human-readable validation report."""
lines = []
lines.append("=" * 70)
lines.append(" PatchJudge Validation Report")
lines.append("=" * 70)
lines.append(f"\n📊 Dataset: {vr.total_examples} examples")
# Score distribution
lines.append(f"\n📈 Score Distribution:")
lines.append(f" Mean: {vr.score_mean:.1f}")
lines.append(f" Median: {vr.score_median:.1f}")
lines.append(f" Std: {vr.score_std:.1f}")
# Score histogram
bins = [0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100]
scores = [r.merge_score for r in results]
hist = defaultdict(int)
for s in scores:
for i in range(len(bins) - 1):
if bins[i] <= s < bins[i+1]:
hist[f"{bins[i]}-{bins[i+1]}"] += 1
break
else:
hist[f"90-100"] += 1
lines.append(f"\n Score Distribution:")
for label in [f"{bins[i]}-{bins[i+1]}" for i in range(len(bins)-1)]:
count = hist.get(label, 0)
bar = "█" * count
lines.append(f" {label:>7}: {bar} ({count})")
# METR alignment
lines.append(f"\n🎯 METR Alignment:")
lines.append(
f" Test-passing patches below {self.merge_threshold}: "
f"{vr.test_passing_below_50_pct:.1%}"
)
metr_target = 0.50
if abs(vr.test_passing_below_50_pct - metr_target) < 0.15:
lines.append(f" ✅ ALIGNED with METR finding (~50% not merge-worthy)")
elif vr.test_passing_below_50_pct < metr_target - 0.15:
lines.append(f" ⚠️ Too lenient — scoring too many patches as merge-worthy")
else:
lines.append(f" ⚠️ Too harsh — scoring too many patches as not merge-worthy")
# Resolved vs Unresolved
lines.append(f"\n🔀 Resolved vs Unresolved Separation:")
lines.append(f" Mean score (resolved): {vr.mean_score_resolved:.1f}")
lines.append(f" Mean score (unresolved): {vr.mean_score_unresolved:.1f}")
lines.append(f" Separation: {vr.mean_score_resolved - vr.mean_score_unresolved:+.1f}")
lines.append(f" Correlation: {vr.score_resolved_correlation:.3f}")
# Known-bad detection
if vr.known_bad_total > 0:
lines.append(f"\n🚨 Known-Bad Pattern Detection:")
lines.append(
f" Detected: {vr.known_bad_detected}/{vr.known_bad_total} "
f"({vr.known_bad_detection_rate:.1%})"
)
if vr.known_bad_detection_rate >= 0.80:
lines.append(f" ✅ Good detection rate")
else:
lines.append(f" ⚠️ Detection rate below 80% — judge may be too lenient")
# Per-dimension stats
lines.append(f"\n📐 Per-Dimension Scores:")
for dim in ["correctness", "completeness", "code_quality",
"non_regression_risk", "merge_readiness"]:
stats = vr.dimension_stats.get(dim, {})
if stats:
lines.append(
f" {dim:>25}: "
f"mean={stats['mean']:.1f} "
f"std={stats['std']:.1f} "
f"[{stats['min']}-{stats['max']}]"
)
# Top flags
all_flags = defaultdict(int)
for r in results:
for dim, data in r.dimension_scores.items():
for flag in data.get("flags", []):
if flag and flag != "JUDGE_ERROR":
all_flags[flag] += 1
if all_flags:
lines.append(f"\n🏴 Most Common Flags:")
for flag, count in sorted(all_flags.items(), key=lambda x: -x[1])[:10]:
lines.append(f" {count:>4}x {flag}")
# Example best/worst
scored = list(zip(examples, results))
scored.sort(key=lambda x: x[1].merge_score, reverse=True)
if len(scored) >= 3:
lines.append(f"\n⭐ Top 3 Patches:")
for ex, r in scored[:3]:
lines.append(
f" {r.merge_score:5.1f} {ex.instance_id} "
f"({ex.agent_name}, {'PASS' if ex.test_passed else 'FAIL'})"
)
lines.append(f"\n💀 Bottom 3 Patches:")
for ex, r in scored[-3:]:
lines.append(
f" {r.merge_score:5.1f} {ex.instance_id} "
f"({ex.agent_name}, {'PASS' if ex.test_passed else 'FAIL'})"
)
lines.append("\n" + "=" * 70)
report = '\n'.join(lines)
return report
def run_full_validation(
examples: list[PatchExample],
results: list[JudgeResult],
gold_data: Optional[list[dict]] = None,
judge=None,
) -> tuple[ValidationResult, str]:
"""Run the complete validation pipeline.
Args:
examples: Judged patch examples.
results: Judge results for those examples.
gold_data: Gold standard data for generating known-bad patches.
judge: PatchJudge instance (needed if judging known-bad patches).
Returns:
(ValidationResult, report_string)
"""
known_bad_results = None
# Generate and judge known-bad patches if we have gold data and a judge
if gold_data and judge:
logger.info("Generating known-bad patches...")
bad_patches = KnownBadPatchGenerator.generate_all(gold_data)
if bad_patches:
logger.info(f"Judging {len(bad_patches)} known-bad patches...")
bad_judge_results = judge.judge_batch(bad_patches, show_progress=True)
known_bad_results = list(zip(bad_patches, bad_judge_results))
# Print known-bad summary
for bp, br in known_bad_results[:5]:
logger.info(
f" Known-bad [{bp.agent_name}] "
f"{bp.instance_id}: {br.merge_score:.1f}/100"
)
# Run validation
validator = PatchJudgeValidator()
vr = validator.validate(examples, results, known_bad_results)
report = validator.print_report(vr, examples, results)
return vr, report
|