File size: 24,253 Bytes
366b543 | 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 | """
Cascade Validation Runner — proves the static cascade on SWE-bench.
This script runs the cascade agent on selected SWE-bench instances and
verifies the patches by applying test_patch and running FAIL_TO_PASS tests.
Strategy:
1. Clone repo, set up conda environment from environment_setup_commit
2. Run cascade agent (T1 Llama-3.1-8B → T2 Llama-3.3-70B)
3. Apply model patch + test_patch
4. Run FAIL_TO_PASS tests via pytest
5. Record: resolved, cost, model tier used, token counts
The critical question this answers:
"Do the 10 cascade-only (T1/T2 solves where both T4 models fail)
instances produce valid patches, or are they weak-test passes?"
Requirements:
- conda (mamba preferred for speed)
- git
- HF_TOKEN (free inference via huggingface_hub)
- No Docker needed
Usage:
python validate_cascade.py --instance django__django-12308
python validate_cascade.py --batch 10 --target cascade-only
python validate_cascade.py --batch 50 --target all
"""
import argparse
import json
import os
import re
import subprocess
import sys
import tempfile
import time
import traceback
from datetime import datetime
from pathlib import Path
from typing import Optional, Tuple, Dict, List
# ============================================================
# SWE-BENCH CONSTANTS
# ============================================================
# The 10 cascade-only instances (T1 or T2 solves, both T4 models fail)
# Extracted from the trace simulation in CORRECTED_REPORT.md
CASCADE_ONLY_INSTANCES = [
"astropy__astropy-14365",
"astropy__astropy-14995",
"django__django-11815",
"django__django-13089",
"django__django-13807",
"django__django-14315",
"matplotlib__matplotlib-25224",
"matplotlib__matplotlib-25311",
"sympy__sympy-19487",
"sympy__sympy-20590",
]
# The full set includes 14 frontier-retry-only instances for comparison
FRONTIER_ONLY_INSTANCES = [
"django__django-12453",
"django__django-14030",
"django__django-14349",
"django__django-14855",
"django__django-15098",
"django__django-16235",
"matplotlib__matplotlib-26020",
"psf__requests-6028",
"pylint-dev__pylint-7080",
"scikit-learn__scikit-learn-13439",
"scikit-learn__scikit-learn-14087",
"sphinx-doc__sphinx-10323",
"sphinx-doc__sphinx-10466",
"sphinx-doc__sphinx-10614",
]
REPO_URLS = {
"django/django": "https://github.com/django/django.git",
"pytest-dev/pytest": "https://github.com/pytest-dev/pytest.git",
"scikit-learn/scikit-learn": "https://github.com/scikit-learn/scikit-learn.git",
"sympy/sympy": "https://github.com/sympy/sympy.git",
"matplotlib/matplotlib": "https://github.com/matplotlib/matplotlib.git",
"sphinx-doc/sphinx": "https://github.com/sphinx-doc/sphinx.git",
"astropy/astropy": "https://github.com/astropy/astropy.git",
"psf/requests": "https://github.com/psf/requests.git",
"pylint-dev/pylint": "https://github.com/pylint-dev/pylint.git",
}
def run(cmd: list, cwd: str = None, timeout: int = 180, env: dict = None) -> Tuple[int, str, str]:
"""Run a command, return (returncode, stdout, stderr)."""
try:
result = subprocess.run(
cmd, cwd=cwd, capture_output=True, text=True,
timeout=timeout, env=env or os.environ
)
return result.returncode, result.stdout, result.stderr
except subprocess.TimeoutExpired:
return 124, "", "TIMEOUT"
except Exception as e:
return -1, "", str(e)
# ============================================================
# CASCADE AGENT (using free HF Inference API)
# ============================================================
def call_hf_model(
model_id: str,
messages: list,
max_tokens: int = 4096,
temperature: float = 0.2
) -> Tuple[str, int, int]:
"""
Call a model via HF Inference API (free).
Returns (response_text, input_tokens, output_tokens).
"""
from huggingface_hub import InferenceClient
client = InferenceClient(model_id)
completion = client.chat.completions.create(
model=model_id,
messages=messages,
max_tokens=max_tokens,
temperature=temperature,
)
response = completion.choices[0].message.content
input_tokens = getattr(completion.usage, "prompt_tokens", 0) if hasattr(completion, "usage") and completion.usage else len(messages) * 100
output_tokens = getattr(completion.usage, "completion_tokens", 0) if hasattr(completion, "usage") and completion.usage else len(response) // 4
return response, input_tokens, output_tokens
def build_cascade_messages(
instance: dict,
repo_dir: Path,
previous_failure: Optional[str] = None,
) -> list:
"""Build the message list for the cascade agent."""
problem = instance.get("problem_statement", "")
hint = instance.get("hints_text", "")
system_prompt = """You are a software engineer fixing a bug in an open-source project.
Your task is to produce a correct patch that fixes the issue described.
You have access to a bash shell in the repository directory. Use it to:
- Explore the codebase (ls, find, grep, git log)
- Read files (cat, head)
- Run existing tests (pytest)
- Edit files with sed or write tools
Output format:
- For bash commands: <bash>command here</bash>
- For your final patch: <patch>diff here</patch>
- When done and tests pass: <submit>Done</submit>
Be thorough. Read the relevant code, understand the bug, make a minimal fix,
and verify it passes the tests."""
# Build the task description
task = f"""Repository: {instance['repo']}
Base commit: {instance['base_commit'][:12]}
PROBLEM:
{problem}"""
if hint:
task += f"\n\nHINT: {hint}"
task += f"""
The repository is at {repo_dir}. Your bash commands will run from that directory.
Start by exploring the codebase to understand the issue, then implement and test your fix."""
if previous_failure:
system_prompt += f"\n\nYour previous attempt failed with the following issues:\n{previous_failure}\nPlease fix your approach."
return [
{"role": "system", "content": system_prompt},
{"role": "user", "content": task},
]
def extract_patch(response: str) -> Optional[str]:
"""Extract a git patch from the agent's response."""
# Try <patch> tags
m = re.search(r'<patch>(.*?)</patch>', response, re.DOTALL)
if m:
return m.group(1).strip()
# Try ```diff blocks
m = re.search(r'```diff\s*\n(.*?)```', response, re.DOTALL)
if m:
return m.group(1).strip()
# Try ```patch blocks
m = re.search(r'```patch\s*\n(.*?)```', response, re.DOTALL)
if m:
return m.group(1).strip()
# Try to find diff content directly
diff_match = re.search(r'(diff --git.*?)(?:\n```|\n<patch>|\n<submit>|\Z)', response, re.DOTALL)
if diff_match:
return diff_match.group(1).strip()
return None
def run_cascade_agent(
instance: dict,
repo_dir: Path,
max_turns: int = 30,
) -> dict:
"""
Run the cascade agent on one instance.
Tiers:
T1: meta-llama/Llama-3.1-8B-Instruct (free, fast, weak)
T2: meta-llama/Llama-3.3-70B-Instruct (free, moderate)
Returns: {patch, resolved, tier, tokens, turns, cost}
"""
TIERS = [
{"name": "T1", "model": "meta-llama/Llama-3.1-8B-Instruct", "max_turns": max_turns, "cost_per_1k": 0.0},
{"name": "T2", "model": "meta-llama/Llama-3.3-70B-Instruct", "max_turns": max_turns, "cost_per_1k": 0.0},
]
result = {
"instance_id": instance["instance_id"],
"patch": None,
"resolved": False,
"tier_used": None,
"total_turns": 0,
"total_input_tokens": 0,
"total_output_tokens": 0,
"error": None,
}
previous_failures = []
for tier in TIERS:
print(f"\n [{tier['name']}] {tier['model']} (max {tier['max_turns']} turns)")
messages = build_cascade_messages(
instance, repo_dir,
previous_failure="\n".join(previous_failures) if previous_failures else None
)
tier_input_tokens = 0
tier_output_tokens = 0
for turn in range(tier['max_turns']):
try:
response, in_tok, out_tok = call_hf_model(
tier['model'], messages,
max_tokens=4096,
temperature=0.2
)
tier_input_tokens += in_tok
tier_output_tokens += out_tok
# Add response to conversation
messages.append({"role": "assistant", "content": response})
print(f" Turn {turn+1}: {in_tok}+{out_tok} tokens")
# Check for patch submission
patch = extract_patch(response)
if patch:
print(f" → Patch found! ({len(patch)} chars)")
result["patch"] = patch
result["tier_used"] = tier["name"]
result["total_turns"] = turn + 1
result["total_input_tokens"] = tier_input_tokens
result["total_output_tokens"] = tier_output_tokens
return result
# Check for bash commands
bash_commands = re.findall(r'<bash>(.*?)</bash>', response, re.DOTALL)
for cmd in bash_commands:
cmd = cmd.strip()
print(f" $ {cmd[:100]}...")
rc, stdout, stderr = run(
["bash", "-c", cmd],
cwd=str(repo_dir),
timeout=60
)
output = stdout.strip()[:2000]
if stderr.strip():
output += f"\n[stderr: {stderr.strip()[:500]}]"
if rc != 0:
output += f"\n[EXIT CODE: {rc}]"
messages.append({"role": "user", "content": f"<output>\n{output}\n</output>"})
# Check for submit
if "<submit>" in response:
print(f" → Submitted but no patch found")
break
except Exception as e:
print(f" Error: {e}")
previous_failures.append(f"[{tier['name']}] turn {turn+1} error: {str(e)[:200]}")
break
previous_failures.append(f"[{tier['name']}] failed after {tier['max_turns']} turns — no patch produced")
result["error"] = "All tiers exhausted without producing a patch"
return result
# ============================================================
# PATCH VERIFICATION
# ============================================================
def verify_cascade_patch(
instance: dict,
model_patch: str,
repo_dir: Path,
env_name: str = None,
) -> dict:
"""
Verify that a model-generated patch passes SWE-bench tests.
1. Reset repo to base_commit
2. Apply model patch
3. Apply test_patch
4. Run FAIL_TO_PASS tests
5. Run PASS_TO_PASS tests (regression check)
"""
result = {
"resolved": False,
"all_f2p_pass": False,
"all_p2p_pass": False,
"f2p_failures": [],
"p2p_failures": [],
"error": None,
}
try:
# Reset to base_commit
base_commit = instance.get("base_commit", "")
if base_commit:
rc, _, _ = run(["git", "checkout", "-f", base_commit], cwd=str(repo_dir))
if rc != 0:
result["error"] = f"could not checkout base_commit {base_commit[:12]}"
return result
# Apply model patch
patch_file = repo_dir / "aco_model.patch"
patch_file.write_text(model_patch)
rc, out, err = run(
["git", "apply", "--check", str(patch_file)],
cwd=str(repo_dir)
)
if rc != 0:
result["error"] = f"patch --check failed: {err[:300]}"
return result
rc, out, err = run(
["git", "apply", str(patch_file)],
cwd=str(repo_dir)
)
if rc != 0:
# Try with --reject
rc, out, err = run(
["git", "apply", "--reject", str(patch_file)],
cwd=str(repo_dir)
)
if rc != 0:
result["error"] = f"patch apply failed: {err[:300]}"
return result
# Apply test_patch
test_patch = instance.get("test_patch", "")
if not test_patch:
result["error"] = "no test_patch in instance"
return result
test_file = repo_dir / "aco_test.patch"
test_file.write_text(test_patch)
rc, out, err = run(
["git", "apply", "--check", str(test_file)],
cwd=str(repo_dir)
)
if rc == 0:
rc, out, err = run(
["git", "apply", str(test_file)],
cwd=str(repo_dir)
)
if rc != 0:
rc, out, err = run(
["git", "apply", "--reject", str(test_file)],
cwd=str(repo_dir)
)
# Run FAIL_TO_PASS tests
f2p = instance.get("FAIL_TO_PASS", [])
if not f2p:
result["error"] = "no FAIL_TO_PASS tests"
return result
print(f" Running {len(f2p)} FAIL_TO_PASS tests...")
cmd_prefix = ["conda", "run", "-n", env_name] if env_name else []
cmd = cmd_prefix + ["python", "-m", "pytest", "-v", "--tb=short", "-x"] + f2p
rc, out, err = run(cmd, cwd=str(repo_dir), timeout=300)
if rc == 0:
result["all_f2p_pass"] = True
# Run PASS_TO_PASS regression tests
p2p = instance.get("PASS_TO_PASS", [])
if p2p:
print(f" Running {len(p2p)} PASS_TO_PASS regression tests...")
cmd2 = cmd_prefix + ["python", "-m", "pytest", "-v", "--tb=short", "-x"] + p2p[:20]
rc2, out2, err2 = run(cmd2, cwd=str(repo_dir), timeout=300)
if rc2 == 0:
result["all_p2p_pass"] = True
result["resolved"] = True
else:
result["error"] = f"P2P regression: {(out2+err2)[:300]}"
result["p2p_failures"] = [l.strip() for l in (out2+err2).split('\n') if 'FAILED' in l and '::' in l]
else:
result["resolved"] = True
else:
result["error"] = f"F2P failures: {(out+err)[:500]}"
result["f2p_failures"] = [l.strip() for l in (out+err).split('\n') if 'FAILED' in l and '::' in l]
return result
except Exception as e:
result["error"] = f"verification error: {str(e)[:300]}"
return result
# ============================================================
# MAIN VALIDATION PIPELINE
# ============================================================
def validate_one(instance: dict) -> dict:
"""
Full validation pipeline for one instance:
1. Clone repo
2. Set up conda environment
3. Run cascade agent
4. Verify patch
5. Report results
"""
inst_id = instance["instance_id"]
repo = instance.get("repo", "")
env_setup_commit = instance.get("environment_setup_commit", "")
base_commit = instance.get("base_commit", "")
result = {
"instance_id": inst_id,
"repo": repo,
"timestamp": datetime.now().isoformat(),
"stages": {},
"final_resolved": False,
"tier_used": None,
"total_cost": 0.0,
"error": None,
}
print(f"\n{'='*70}")
print(f"VALIDATING: {inst_id}")
print(f" Repo: {repo} Base: {base_commit[:12]} EnvSetup: {env_setup_commit[:12]}")
print(f"{'='*70}")
with tempfile.TemporaryDirectory(prefix=f"aco_valid_{inst_id.replace('/', '_')}_") as tmpdir:
work_dir = Path(tmpdir)
repo_dir = work_dir / "repo"
env_name = f"aco_{inst_id.replace('__', '_').replace('-', '_')[:40]}"
# STAGE 1: Clone repo
print("\n--- Stage 1: Clone repo ---")
repo_url = REPO_URLS.get(repo, f"https://github.com/{repo}.git")
t0 = time.time()
rc, out, err = run(["git", "clone", "--depth", "50", repo_url, str(repo_dir)], timeout=300)
if rc != 0:
# Retry without depth limit
rc, out, err = run(["git", "clone", repo_url, str(repo_dir)], timeout=600)
result["stages"]["clone"] = {"success": rc == 0, "duration": time.time() - t0}
if rc != 0:
result["error"] = f"clone failed: {err[:300]}"
return result
# STAGE 2: Set up environment
print("\n--- Stage 2: Set up conda environment ---")
t0 = time.time()
# Checkout env_setup_commit to find environment.yml
if env_setup_commit:
run(["git", "fetch", "origin", env_setup_commit], cwd=str(repo_dir), timeout=60)
run(["git", "checkout", env_setup_commit], cwd=str(repo_dir), timeout=30)
# Find environment.yml
env_candidates = [
"environment.yml", "dev/environment.yml", ".github/environment.yml",
"ci/environment.yml", ".azure-pipelines/environment.yml",
]
env_yml = None
for c in env_candidates:
p = repo_dir / c
if p.exists():
env_yml = p
break
if not env_yml:
for p in repo_dir.rglob("environment.yml"):
if p.stat().st_size > 10:
env_yml = p
break
env_setup_ok = False
if env_yml:
print(f" Using environment.yml: {env_yml.relative_to(repo_dir)}")
rc, out, err = run(
["conda", "env", "create", "-f", str(env_yml), "-n", env_name, "--quiet"],
timeout=600
)
env_setup_ok = (rc == 0)
else:
print(f" No environment.yml found, creating python=3.10 env")
rc, out, err = run(
["conda", "create", "-n", env_name, "python=3.10", "pip", "-y", "--quiet"],
timeout=300
)
env_setup_ok = (rc == 0)
if not env_setup_ok:
result["stages"]["environment"] = {"success": False, "error": err[:300], "duration": time.time() - t0}
result["error"] = f"env setup failed: {err[:300]}"
return result
# Checkout base_commit and install
if base_commit:
run(["git", "fetch", "origin", base_commit], cwd=str(repo_dir), timeout=60)
run(["git", "checkout", base_commit], cwd=str(repo_dir), timeout=30)
run(["conda", "run", "-n", env_name, "pip", "install", "-e", ".", "--quiet"],
cwd=str(repo_dir), timeout=300)
result["stages"]["environment"] = {"success": True, "duration": time.time() - t0}
print(f" Environment ready in {time.time() - t0:.1f}s")
# STAGE 3: Run cascade agent
print("\n--- Stage 3: Run cascade agent ---")
t0 = time.time()
agent_result = run_cascade_agent(instance, repo_dir, max_turns=30)
result["stages"]["agent"] = {
"success": agent_result["patch"] is not None,
"tier": agent_result["tier_used"],
"turns": agent_result["total_turns"],
"input_tokens": agent_result["total_input_tokens"],
"output_tokens": agent_result["total_output_tokens"],
"duration": time.time() - t0,
}
if not agent_result["patch"]:
result["error"] = "No patch produced by any tier"
return result
# STAGE 4: Verify patch
print("\n--- Stage 4: Verify patch ---")
t0 = time.time()
verify_result = verify_cascade_patch(instance, agent_result["patch"], repo_dir, env_name)
result["stages"]["verify"] = {
"resolved": verify_result["resolved"],
"all_f2p_pass": verify_result.get("all_f2p_pass", False),
"all_p2p_pass": verify_result.get("all_p2p_pass", False),
"error": verify_result.get("error"),
"duration": time.time() - t0,
}
result["final_resolved"] = verify_result["resolved"]
result["tier_used"] = agent_result["tier_used"]
if verify_result.get("error"):
result["error"] = verify_result["error"]
# Cleanup
print(f"\n Cleaning up conda env {env_name}...")
run(["conda", "env", "remove", "-n", env_name, "-y", "--quiet"], timeout=30)
return result
def main():
parser = argparse.ArgumentParser(description="Cascade Validation Runner")
parser.add_argument("--instance", type=str, help="Single instance ID")
parser.add_argument("--batch", type=int, default=10, help="Number of instances to validate")
parser.add_argument("--target", choices=["cascade-only", "frontier-only", "all"], default="cascade-only")
parser.add_argument("--output", type=str, default="validation_results.jsonl")
args = parser.parse_args()
from datasets import load_dataset
print("Loading SWE-bench_Verified...")
ds = load_dataset("princeton-nlp/SWE-bench_Verified", split="test")
# Select instances
if args.instance:
instances = [dict(row) for row in ds if row["instance_id"] == args.instance]
if not instances:
print(f"Instance {args.instance} not found")
sys.exit(1)
elif args.target == "cascade-only":
instances = [dict(row) for row in ds if row["instance_id"] in CASCADE_ONLY_INSTANCES]
instances = instances[:args.batch]
elif args.target == "frontier-only":
instances = [dict(row) for row in ds if row["instance_id"] in FRONTIER_ONLY_INSTANCES]
instances = instances[:args.batch]
else:
instances = [dict(row) for row in ds][:args.batch]
print(f"Selected {len(instances)} instances for validation\n")
results = []
for i, instance in enumerate(instances):
print(f"\n{'#'*70}")
print(f" [{i+1}/{len(instances)}] {instance['instance_id']}")
print(f"{'#'*70}")
try:
result = validate_one(instance)
results.append(result)
# Print summary
status = "✅ RESOLVED" if result["final_resolved"] else "❌ FAILED"
print(f"\n {status} | Tier: {result['tier_used']} | error: {result.get('error', 'none')}")
except Exception as e:
print(f"\n ❌ CRASH: {e}")
traceback.print_exc()
results.append({
"instance_id": instance["instance_id"],
"final_resolved": False,
"error": str(e),
})
# Save incrementally
with open(args.output, "w") as f:
for r in results:
f.write(json.dumps(r) + "\n")
print(f"\n ← Saved to {args.output} ({len(results)} results so far)")
# Final report
resolved = [r for r in results if r["final_resolved"]]
t1_resolved = [r for r in resolved if r.get("tier_used") == "T1"]
t2_resolved = [r for r in resolved if r.get("tier_used") == "T2"]
print(f"\n{'='*70}")
print(f"VALIDATION COMPLETE")
print(f"{'='*70}")
print(f" Total: {len(results)}")
print(f" Resolved: {len(resolved)} ({len(resolved)/max(len(results),1)*100:.1f}%)")
print(f" T1 (Llama-3.1-8B): {len(t1_resolved)}")
print(f" T2 (Llama-3.3-70B): {len(t2_resolved)}")
print(f" Failed: {len(results) - len(resolved)}")
print(f" Results: {args.output}")
if __name__ == "__main__":
main()
|