File size: 11,768 Bytes
29c4a80 | 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 | """
Batch Cascade Validation Script
Runs the cascade agent on MULTIPLE SWE-bench instances and
verifies each patch via conda environment + pytest.
This is the script that proves the cascade causally, not just
correlationally from trace simulation.
Usage:
python batch_validate.py --instances 3 --target cascade-only
python batch_validate.py --instances 5 --target all
Requirements: hf_jobs with a10g-largex2, 8h timeout
"""
import json, os, re, subprocess, sys, tempfile, time, traceback
from datetime import datetime
from pathlib import Path
# ============================================================
# INSTANCE LISTS
# ============================================================
CASCADE_ONLY = [
"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",
]
FRONTIER_ONLY = [
"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",
]
def sh(cmd, cwd=None, timeout=120):
r = subprocess.run(cmd, cwd=cwd, capture_output=True, text=True, timeout=timeout, shell=True)
return r.returncode, r.stdout, r.stderr
def ensure_conda():
for p in [os.path.expanduser("~/miniconda3/bin/conda"), "/opt/conda/bin/conda"]:
if os.path.exists(p): return p
print("π¦ Installing Miniconda...")
sh("wget -q https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O /tmp/miniconda.sh && bash /tmp/miniconda.sh -b -p $HOME/miniconda3", timeout=300)
p = os.path.expanduser("~/miniconda3/bin/conda")
sh(f"{p} config --set always_yes yes --set changeps1 no && {p} tos accept --override-channels --channel https://repo.anaconda.com/pkgs/main 2>/dev/null; true", timeout=30)
os.environ["PATH"] = os.path.expanduser("~/miniconda3/bin:") + os.environ.get("PATH", "")
return p
def is_valid_patch(text):
return bool(text) and len(text) > 10 and 'diff --git' in text and '@@' in text
def extract_patch(text):
m = re.search(r'<patch>\s*\n?(.*?)</patch>', text, re.DOTALL)
if m and is_valid_patch(m.group(1)): return m.group(1).strip()
m = re.search(r'```diff\s*\n(.*?)```', text, re.DOTALL)
if m and is_valid_patch(m.group(1)): return m.group(1).strip()
di = text.find('diff --git')
if di >= 0:
patch = text[di:].strip()[:2000]
if is_valid_patch(patch): return patch
return None
def call_model(client, messages, max_tokens=4096):
try:
c = client.chat.completions.create(model=client.model, messages=messages, max_tokens=max_tokens, temperature=0.2)
t = c.choices[0].message.content
it = c.usage.prompt_tokens if hasattr(c,'usage') and c.usage else 0
ot = c.usage.completion_tokens if hasattr(c,'usage') and c.usage else len(t)//4
return t, it, ot
except Exception as e: return f"[ERROR: {e}]", 0, 0
def run_cascade(instance, repo_dir, conda, env_name):
from huggingface_hub import InferenceClient
T1, T2 = "meta-llama/Llama-3.1-8B-Instruct", "meta-llama/Llama-3.3-70B-Instruct"
system = f"Fix bug in {instance['repo']}. Repo: {repo_dir}.\n<bash>cmd</bash>\n<patch>\ndiff --git a/path b/path\n--- a/path\n+++ b/path\n@@ -N,M +N,M @@\nchanges\n</patch>\n<submit>Done</submit>"
messages = [{"role":"system","content":system},{"role":"user","content":f"PROBLEM:\n{instance.get('problem_statement','')}\n\nExplore and fix."}]
for tier, mid, mt in [("T1",T1,30),("T2",T2,30)]:
print(f"\n[{tier}] {mid}")
client = InferenceClient(mid)
ti = to = 0
for turn in range(mt):
text, it, ot = call_model(client, messages, 4096)
ti += it; to += ot
messages.append({"role":"assistant","content":text})
print(f" T{turn+1}: {it}+{ot} tok")
patch = extract_patch(text)
if patch:
(Path(repo_dir)/"_c.patch").write_text(patch)
rc, out, err = sh(f"cd {repo_dir} && git apply --check _c.patch 2>&1", timeout=10)
if rc == 0:
print(f" β
VALID ({len(patch)}ch)")
return {"patch":patch,"tier":tier,"turns":turn+1,"input_tokens":ti,"output_tokens":to}
print(f" β Invalid: {err[:100]}")
messages.append({"role":"user","content":f"Patch check failed: {err[:200]}\nUse git diff for valid unified diff."})
continue
for cmd in re.findall(r'<bash>(.*?)</bash>', text, re.DOTALL):
cmd = cmd.strip().replace("pytest", f"{conda} run -n {env_name} python -m pytest")
rc, out, err = sh(cmd, cwd=str(repo_dir), timeout=60)
o = f"<output>\n{(out+err)[:1500]}\n</output>"
if rc: o = o[:-9] + f" [EXIT:{rc}]\n</output>"
messages.append({"role":"user","content":o})
if "<submit>" in text: break
return {"patch":None,"tier":None,"turns":0,"input_tokens":0,"output_tokens":0}
def verify_patch(instance, patch, repo_dir, conda, env_name):
base = instance.get("base_commit","")
tp = instance.get("test_patch","")
f2p = instance.get("FAIL_TO_PASS",[])
sh(f"cd {repo_dir} && git checkout -f {base} && git clean -fd", timeout=30)
(Path(repo_dir)/"_aco.patch").write_text(patch)
rc, out, err = sh(f"cd {repo_dir} && git apply --check _aco.patch", timeout=10)
if rc: return {"resolved":False,"error":f"patch check: {err[:150]}"}
sh(f"cd {repo_dir} && git apply _aco.patch", timeout=10)
(Path(repo_dir)/"_t.patch").write_text(tp)
sh(f"cd {repo_dir} && (git apply _t.patch) || git apply --reject _t.patch 2>/dev/null; true", timeout=10)
cmd = f"cd {repo_dir} && {conda} run -n {env_name} python -m pytest -v --tb=short -x {' '.join(f2p[:10])}"
rc, out, err = sh(cmd, timeout=300)
if rc == 0:
p2p = instance.get("PASS_TO_PASS",[])
if p2p:
cmd2 = f"cd {repo_dir} && {conda} run -n {env_name} python -m pytest -v --tb=short -x {' '.join(p2p[:10])}"
rc2, out2, err2 = sh(cmd2, timeout=300)
if rc2: return {"resolved":False,"error":f"P2P: {(out2+err2)[:200]}"}
return {"resolved":True,"test_output":(out+err)[:500]}
return {"resolved":False,"error":f"{len(re.findall(r'FAILED', out+err))} F2P failures","test_output":(out+err)[:500]}
def setup_env(conda, instance, repo_dir, env_name):
ec = instance.get("environment_setup_commit","")
if ec:
sh(f"cd {repo_dir} && git fetch origin {ec} && git checkout {ec}", timeout=60)
eyml = None
for c in ["environment.yml","dev/environment.yml",".github/environment.yml","ci/environment.yml"]:
if (Path(repo_dir)/c).exists(): eyml = c; break
if eyml:
rc, out, err = sh(f"{conda} env create -f {repo_dir}/{eyml} -n {env_name} --quiet", timeout=600)
else:
rc, out, err = sh(f"{conda} create -n {env_name} python=3.10 pip -y", timeout=300)
if rc:
rc, out, err = sh(f"{conda} create -n {env_name} python=3.10 pip -y", timeout=300)
if rc: return False, f"conda: {err[:200]}"
base = instance["base_commit"]
sh(f"cd {repo_dir} && git fetch origin {base} && git checkout {base}", timeout=60)
sh(f"cd {repo_dir} && {conda} run -n {env_name} pip install -e . 2>&1 | tail -3", timeout=300)
sh(f"cd {repo_dir} && {conda} run -n {env_name} pip install . 2>&1 | tail -3", timeout=300)
return True, ""
def main():
import datasets
target = os.environ.get("INSTANCE_TARGET", "cascade-only")
max_instances = int(os.environ.get("MAX_INSTANCES", "3"))
print(f"π BATCH CASCADE VALIDATION β target={target} max={max_instances}")
print(f" {datetime.now().isoformat()}")
conda = ensure_conda()
if not conda: print("β No conda"); sys.exit(1)
ds = datasets.load_dataset("princeton-nlp/SWE-bench_Verified", split="test")
if target == "cascade-only":
iids = CASCADE_ONLY[:max_instances]
elif target == "frontier-only":
iids = FRONTIER_ONLY[:max_instances]
else:
iids = [r["instance_id"] for r in ds][:max_instances]
instances = {}
for row in ds:
if row["instance_id"] in iids:
instances[row["instance_id"]] = dict(row)
print(f"Instances: {iids}\n")
results = []
for i, iid in enumerate(iids):
instance = instances.get(iid)
if not instance: continue
print(f"\n{'='*60}\n[{i+1}/{len(iids)}] {iid}\n{'='*60}")
with tempfile.TemporaryDirectory(prefix=f"aco_{i}_") as tmpdir:
repo_dir = Path(tmpdir) / "repo"
env_name = f"aco_{iid.replace('__','_').replace('-','_')[:30]}"
print(f"Clone...")
url = f"https://github.com/{instance['repo']}.git"
rc, out, err = sh(f"git clone --depth 100 {url} {repo_dir}", timeout=600)
if rc:
results.append({"instance_id":iid,"resolved":False,"error":f"Clone: {err[:200]}"})
continue
print(f"Env...")
ok, err = setup_env(conda, instance, repo_dir, env_name)
if not ok:
results.append({"instance_id":iid,"resolved":False,"error":f"Env: {err}"})
continue
print(f"Cascade...")
agent = run_cascade(instance, repo_dir, conda, env_name)
if not agent["patch"]:
results.append({"instance_id":iid,"resolved":False,"tier":None,"error":"No valid patch"})
sh(f"{conda} env remove -n {env_name} -y --quiet 2>/dev/null; true", timeout=30)
continue
print(f"Verify...")
verify = verify_patch(instance, agent["patch"], repo_dir, conda, env_name)
r = {
"instance_id":iid, "repo":instance["repo"],
"resolved":verify["resolved"], "tier":agent["tier"],
"turns":agent["turns"], "input_tokens":agent["input_tokens"],
"output_tokens":agent["output_tokens"],
"error":verify.get("error"),
"timestamp":datetime.now().isoformat()
}
results.append(r)
status = "β
" if verify["resolved"] else "β"
print(f" {status} {agent['tier']} {agent['turns']}t")
sh(f"{conda} env remove -n {env_name} -y --quiet 2>/dev/null; true", timeout=30)
# Incremental save
with open("batch_results.jsonl","w") as f:
for r in results: f.write(json.dumps(r)+"\n")
resolved = [r for r in results if r["resolved"]]
t1 = [r for r in resolved if r.get("tier")=="T1"]
t2 = [r for r in resolved if r.get("tier")=="T2"]
print(f"\n{'='*60}\nRESULTS: {len(resolved)}/{len(results)} resolved")
print(f" T1: {len(t1)} T2: {len(t2)}")
for r in results:
s = "β
" if r["resolved"] else "β"
print(f" {s} {r['instance_id']} [{r.get('tier','')}]")
print(f"Saved: batch_results.jsonl")
if __name__=="__main__":
try: sys.exit(main())
except Exception as e: print(f"π₯ {e}"); traceback.print_exc(); sys.exit(1)
|