Spaces:
Running
Running
File size: 11,657 Bytes
f791e67 | 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 | #!/usr/bin/env python3
"""Independent public-artifact audit for the exact Table 5 and Table 6 claims."""
from __future__ import annotations
import hashlib
import io
import json
import math
import subprocess
import sys
import tarfile
from pathlib import Path
import requests
ROOT = Path(__file__).resolve().parents[2]
ARTIFACT_DIR = ROOT / ".openresearch/artifacts/claim5_ablation_scaling"
ROUTES = ARTIFACT_DIR / "four_routes.json"
EXACT_GATE = ROOT / "repro/src/claim5_exact_gate.py"
USER_AGENT = "OpenResearch-Reproduction/1.0 paper-2509.26476"
REGRESS_REPO = "google-deepmind/regress-lm"
REGRESS_REVISIONS = {
"paper_time": "b36c45898c88f96bd9e3bd73a7d17c89f9d73f0b",
"current_audited": "6c23ccb51ae9862d98af9faffc1286ca2149b12e",
}
QIN_REPO = "shiwenqin/transferrable-surrogates"
QIN_REVISION = "d087a70ec483fb8cee7536c99a0f9c363609eb05"
QIN_PATH = "lm_tuning/src/fine_tuning/bert_loo_tuning.py"
QIN_SHA256 = "e5b2e3d31177fe2fa9572cbf99bcf87ff06e5786d2f029a251734f7d7e7558f3"
BASE_MODELS = {
"google/t5gemma-s-s-prefixlm": {
"revision": "7b875178ee4c8a97a5df548972b10cfe200ee695",
"parameters": 312_517_632,
},
"google/t5gemma-b-b-prefixlm": {
"revision": "df5e1422c2a53948a57901c54bfcf397c92bfa1d",
"parameters": 591_490_560,
},
}
TABLE5 = {
"standard_regression_head": 0.478,
"normalized_regression_head": 0.717,
"decoder_head": 0.800,
}
TABLE6 = {"t5gemma_s_s_300m": 0.744, "t5gemma_b_b_600m": 0.782}
EXACT_MARKERS = (
"standard regression head",
"normalized regression head",
"automodelforsequenceclassification",
"minmaxscaler",
"t5gemma-b-b",
"0.717",
"0.782",
)
def get_json(url: str, **params: object) -> tuple[dict | list, bytes]:
response = requests.get(
url,
params=params or None,
headers={"User-Agent": USER_AGENT},
timeout=120,
)
response.raise_for_status()
return response.json(), response.content
def github_tree_audit(revision: str) -> dict:
archive_url = (
f"https://codeload.github.com/{REGRESS_REPO}/tar.gz/{revision}"
)
response = requests.get(
archive_url,
headers={"User-Agent": USER_AGENT},
timeout=120,
)
response.raise_for_status()
archive = tarfile.open(fileobj=io.BytesIO(response.content), mode="r:gz")
members = [member for member in archive.getmembers() if member.isfile()]
paths = sorted(member.name.split("/", 1)[1] for member in members)
dedicated_paths = [
path for path in paths
if any(marker in path.lower() for marker in (
"table5", "table_5", "table6", "table_6", "ablation",
"regression_head", "regression-head", "scaling", "600m",
))
]
marker_hits: dict[str, list[str]] = {marker: [] for marker in EXACT_MARKERS}
inspected_text_files = 0
for member in members:
path = member.name.split("/", 1)[1]
if not path.endswith(
(".py", ".md", ".toml", ".yaml", ".yml", ".json", ".txt", ".sh")
):
continue
extracted = archive.extractfile(member)
if extracted is None:
raise AssertionError(f"archive file could not be read: {path}")
text = extracted.read().decode("utf-8", errors="replace").lower()
inspected_text_files += 1
for marker in EXACT_MARKERS:
if marker in text:
marker_hits[marker].append(path)
exact_hits = {key: value for key, value in marker_hits.items() if value}
return {
"revision": revision,
"archive_url": archive_url,
"archive_sha256": hashlib.sha256(response.content).hexdigest(),
"file_count": len(paths),
"inspected_text_files": inspected_text_files,
"dedicated_table_or_ablation_paths": dedicated_paths,
"exact_marker_hits": exact_hits,
"exact_table_artifact_discovered": bool(dedicated_paths or exact_hits),
}
def author_model_audit() -> dict:
listing, raw = get_json(
"https://huggingface.co/api/models",
author="akhauriyash",
limit=1000,
full="true",
)
if not isinstance(listing, list):
raise AssertionError("unexpected Hugging Face author listing")
ids = sorted(item["id"] for item in listing)
relevant = [
model_id for model_id in ids
if any(token in model_id.lower() for token in ("regress", "rlm", "gemma"))
]
table5_or_600m = [
model_id for model_id in relevant
if any(token in model_id.lower() for token in (
"600m", "b-b", "table5", "ablation", "standard", "normalized",
))
]
return {
"api_url": (
"https://huggingface.co/api/models?author=akhauriyash"
"&limit=1000&full=true"
),
"listing_sha256": hashlib.sha256(raw).hexdigest(),
"relevant_public_models": relevant,
"table5_or_600m_rlm_models": table5_or_600m,
"exact_claim_checkpoint_discovered": bool(table5_or_600m),
}
def base_model_audit() -> dict:
output = {}
for model_id, expected in BASE_MODELS.items():
detail, raw = get_json(f"https://huggingface.co/api/models/{model_id}")
if not isinstance(detail, dict):
raise AssertionError(f"unexpected model response: {model_id}")
if detail.get("sha") != expected["revision"]:
raise AssertionError(f"{model_id}: revision changed")
total = detail.get("safetensors", {}).get("total")
if total != expected["parameters"]:
raise AssertionError(f"{model_id}: parameter count changed")
config_response = requests.get(
f"https://huggingface.co/{model_id}/resolve/{expected['revision']}/config.json",
headers={"User-Agent": USER_AGENT},
timeout=120,
)
output[model_id] = {
"revision": detail["sha"],
"gated": detail.get("gated"),
"safetensors_parameter_count": total,
"metadata_sha256": hashlib.sha256(raw).hexdigest(),
"unauthenticated_config_http_status": config_response.status_code,
"config_publicly_downloadable_without_acceptance": (
config_response.status_code == 200
),
}
return output
def qin_reference_audit() -> dict:
url = (
f"https://raw.githubusercontent.com/{QIN_REPO}/{QIN_REVISION}/{QIN_PATH}"
)
response = requests.get(
url, headers={"User-Agent": USER_AGENT}, timeout=120
)
response.raise_for_status()
digest = hashlib.sha256(response.content).hexdigest()
if digest != QIN_SHA256:
raise AssertionError("pinned Qin reference implementation changed")
text = response.text
required = {
"sequence_classification_regressor": (
"AutoModelForSequenceClassification.from_pretrained" in text
),
"one_output_label": "num_labels=1" in text,
"regression_problem_type": 'problem_type="regression"' in text,
"per_dataset_minmax_0_1": (
"MinMaxScaler(feature_range=(0, 1))" in text
),
}
if not all(required.values()):
raise AssertionError("cited normalized-regression implementation changed")
return {
"url": url,
"sha256": digest,
"features": required,
"scope": (
"Confirms the cited normalized target transformation and scalar "
"sequence-classification regressor pattern; it is not the paper's "
"missing Table 5 implementation, split, or checkpoint."
),
}
def detector_negative_control() -> dict:
synthetic_paths = ["configs/table5_regression_head.yaml"]
detected = any(
any(marker in path.lower() for marker in (
"table5", "regression_head", "ablation", "600m"
))
for path in synthetic_paths
)
if not detected:
raise AssertionError("artifact detector missed synthetic exact artifact")
return {
"synthetic_paths": synthetic_paths,
"expected_detection": True,
"observed_detection": detected,
"passes_claim_acceptance": False,
"reason": "Synthetic file only tests detector sensitivity; it is not evidence.",
}
def verify() -> dict:
repository = {
label: github_tree_audit(revision)
for label, revision in REGRESS_REVISIONS.items()
}
if any(row["exact_table_artifact_discovered"] for row in repository.values()):
raise AssertionError("manual review required: exact marker appeared upstream")
author_models = author_model_audit()
if author_models["exact_claim_checkpoint_discovered"]:
raise AssertionError("manual review required: candidate Claim 5 model appeared")
bases = base_model_audit()
qin = qin_reference_audit()
control = detector_negative_control()
routes = json.loads(ROUTES.read_text())
route_rows = routes["routes"]
if [row["route"] for row in route_rows] != [1, 2, 3, 4]:
raise AssertionError("four-route sequence is incomplete")
if len({row["name"] for row in route_rows}) != 4:
raise AssertionError("routes are not materially distinct")
if route_rows[-1]["name"] != "Dedicated exact-claim falsification attempt":
raise AssertionError("fourth route is not falsification-dedicated")
if any(row["outcome"] != "BLOCKED" for row in route_rows):
raise AssertionError("unsupported route outcome")
if routes["confidence_after_three_routes"] != "LOW":
raise AssertionError("mandatory fourth-route trigger not recorded")
if routes["final_verdict"] != "BLOCKED":
raise AssertionError("public audit cannot decide the exact claim")
gate = subprocess.run(
[sys.executable, str(EXACT_GATE)],
cwd=ROOT,
text=True,
capture_output=True,
check=False,
)
if gate.returncode == 0 or "CLAIM5_EXACT_GATE_BLOCKED" not in gate.stdout:
raise AssertionError("exact claim gate did not fail closed")
improvement = TABLE6["t5gemma_b_b_600m"] - TABLE6["t5gemma_s_s_300m"]
ratio = (
BASE_MODELS["google/t5gemma-b-b-prefixlm"]["parameters"]
/ BASE_MODELS["google/t5gemma-s-s-prefixlm"]["parameters"]
)
if not math.isclose(improvement, 0.038, abs_tol=1e-12):
raise AssertionError("Table 6 arithmetic mismatch")
return {
"status": "PASS",
"claim_verdict": "BLOCKED",
"confidence": "LOW",
"exact_claim_tested": {
"table5": TABLE5,
"table6": TABLE6,
"table6_difference": improvement,
},
"official_repository_snapshots": repository,
"author_public_models": author_models,
"base_models": bases,
"base_parameter_ratio": ratio,
"cited_normalized_head_reference": qin,
"negative_control": control,
"routes_completed": len(route_rows),
"fourth_route": "falsification attempted; no valid counterexample",
"exact_claim_gate": {
"exit_code": gate.returncode,
"stdout": gate.stdout.strip(),
},
"unblockers": routes["unblockers"],
"limitations": [
"Absence from audited public releases does not prove a private artifact does not exist.",
"The public base-model sizes validate rounded 300M/600M labels, not the claimed correlation improvement.",
"The released 181.5M RLM is not identified as either exact Table 6 checkpoint and is not a counterexample.",
],
}
if __name__ == "__main__":
print(json.dumps(verify(), sort_keys=True))
|