Datasets:
Formats:
csv
Sub-tasks:
multi-class-classification
Languages:
English
Size:
10K - 100K
ArXiv:
Tags:
multi-agent-systems
llm-routing
cost-aware-inference
calibration
agent-collaboration
reasoning
License:
File size: 21,958 Bytes
1414c64 2706e3f 1414c64 2706e3f 5c5c8b4 1414c64 5c5c8b4 1414c64 5c5c8b4 1414c64 2706e3f 1414c64 2706e3f 1414c64 2706e3f 1414c64 | 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 | #!/usr/bin/env python3
"""Validate a downloaded copy of the Cost-Aware Protocol Routing dataset.
Usage
-----
python validate.py # validate the directory this file is in
python validate.py --dir PATH # validate another directory
python validate.py --skip-checksums
Exits 0 if every check passes, 1 on a validation failure, 2 on a usage error.
Requires ONLY the Python standard library -- no pandas, no numpy, nothing to
install. It is meant to be the first thing you run after `hf download`.
"""
from __future__ import annotations
import argparse
import csv
import hashlib
import json
import os
import sys
EXPECTED_RELEASE_VERSION = "1.0.0"
EXPECTED_SCHEMA_VERSION = "emnlp-protocol-routing-dataset-v1.0.0"
PROTOCOLS = ["baseline_llm", "single_agent", "per", "broadcast"]
ORACLE_LABELS = PROTOCOLS + ["none"]
REQUIRED_FILES = [
"README.md", "LICENSE", "NOTICE", "CITATION.cff",
"registry/experiments.csv", "registry/benchmarks.csv", "registry/models.csv",
"registry/protocols.csv", "registry/schema.json",
"registry/release_manifest.json", "registry/checksums.sha256",
"data/matched_labels.csv", "data/problems.csv", "data/labels_for_scoring.csv",
"data/probe_inputs.jsonl",
"data/splits/six_setting_splits.csv", "data/splits/primary_omnimath_splits.csv",
"data/router/six_setting_test_predictions.csv", "data/router/six_setting_metrics.csv",
"data/confidence/six_setting_confidence_metrics.csv",
"data/confidence/primary_omnimath_confidence_predictions.csv",
"data/confidence/postanswer_confidence_predictions.csv",
"data/confidence/failure_and_protocol_value_targets.csv",
"data/costs/omnimath_per_protocol_costs.csv",
"registry/example_id_crosswalk.csv",
"docs/confidence_probe_prompt.txt", "docs/primary_confidence_probe_prompt.txt",
"TODO.md", "validate.py",
"data/aggregate/matched_protocol_coverage.csv",
"data/aggregate/oracle_label_distribution.csv",
"data/aggregate/main_routing_heldout.csv",
"docs/schema.md", "docs/provenance.md", "docs/license_audit.md",
"docs/reconstruction.md", "docs/anonymization_report.md",
]
# Optional artifacts: absent by design, listed so their absence is explicit.
OPTIONAL_ARTIFACTS = {
"data/problem_text.csv":
"upstream problem text (withheld for licensing; see docs/reconstruction.md)",
"data/baseline_final_answers.csv":
"Baseline final-answer text (NOT released; probe is not re-runnable without it)",
"data/traces/":
"raw protocol traces / model generations (out of scope for this table-first release)",
"data/costs/other_settings_per_protocol_costs.csv":
"per-problem costs for the other 9 settings (UNRESOLVED -- must not be "
"back-filled from an incompatible token convention; see TODO.md)",
}
EXPECTED_ROWS = {
"data/matched_labels.csv": 15088,
"data/problems.csv": 6803,
"data/labels_for_scoring.csv": 12928,
"data/probe_inputs.jsonl": 12928,
"data/splits/six_setting_splits.csv": 12928,
"data/splits/primary_omnimath_splits.csv": 4181,
"data/router/six_setting_test_predictions.csv": 5832,
"data/confidence/primary_omnimath_confidence_predictions.csv": 839,
"data/confidence/postanswer_confidence_predictions.csv": 12928,
"registry/example_id_crosswalk.csv": 12928,
"data/costs/omnimath_per_protocol_costs.csv": 4155,
}
EXPECTED_SETTINGS = {
"omnimath2__competition_math_4181__gpt_oss_120b": 4181,
"omnimath2__competition_math_4181__gemma_4_31b": 4181,
"labbench__text_no_tool__gpt_oss_120b_text_no_tool": 1542,
"labbench__text_no_tool__gemma_4_31b_text_no_tool": 1542,
"labbench__llm_strict__gpt_oss_120b": 741,
"labbench__llm_strict__gemma_4_31b": 741,
"scibench__text_only__gpt_oss_120b": 565,
"scibench__text_only__gemma_4_31b": 565,
"jeebench__text_only__gpt_oss_120b": 515,
"jeebench__text_only__gemma_4_31b": 515,
}
EXPECTED_BENCHMARKS = {"omnimath2": 4181, "labbench": 1542, "scibench": 565, "jeebench": 515}
EXPECTED_MODELS = {"gpt_oss_120b": 5, "gemma_4_31b": 5}
# Files that must never contain outcome/label columns.
FEATURE_FILES = [
"data/problems.csv", "data/probe_inputs.jsonl",
"data/splits/six_setting_splits.csv", "data/splits/primary_omnimath_splits.csv",
"data/confidence/primary_omnimath_confidence_predictions.csv",
"data/confidence/postanswer_confidence_predictions.csv",
"data/costs/omnimath_per_protocol_costs.csv",
]
FORBIDDEN_IN_FEATURES = {
"oracle_label", "baseline_correct", "single_correct", "per_correct",
"broadcast_correct", "any_protocol_solved", "answer", "gold", "gold_answer",
"problem_text", "question", "reference_solution", "options",
"baseline_final_answer", "baseline_final_passed", "cheapest_successful_protocol",
}
# Infrastructure that download tools and version control create alongside the
# release. Not release content, so never expected in checksums.sha256. Matched by
# PATH COMPONENT, so any depth is covered (e.g. .cache/huggingface/download/*.lock).
IGNORED_DIR_COMPONENTS = {".cache", ".git", "__pycache__", ".ipynb_checkpoints"}
IGNORED_BASENAMES = {".gitattributes", ".gitignore", ".DS_Store", ".gitmodules", "Thumbs.db"}
MAX_LISTED_OFFENDERS = 10
FAILS: list[str] = []
WARNS: list[str] = []
def is_infrastructure(rel_path: str) -> bool:
"""True for tool-created files that are not part of the release payload."""
parts = rel_path.replace(os.sep, "/").split("/")
if any(part in IGNORED_DIR_COMPONENTS for part in parts[:-1]):
return True
return parts[-1] in IGNORED_BASENAMES
def abbreviate(items: "list[str]", limit: int = MAX_LISTED_OFFENDERS) -> str:
"""Render a list without letting it swamp the output."""
shown = ", ".join(items[:limit])
if len(items) > limit:
shown += f", ... and {len(items) - limit} more"
return shown
def ok(label: str, passed: bool, detail: str = "", fail_detail: str = "") -> bool:
"""Report one check.
`detail` is a factual measurement and prints either way. `fail_detail`
explains what went wrong and prints ONLY on failure -- a check must never
print a reason that contradicts its own verdict.
"""
shown = detail if passed else (fail_detail or detail)
print(f" [{'PASS' if passed else 'FAIL'}] {label}" + (f" -- {shown}" if shown else ""))
if not passed:
FAILS.append(f"{label}: {shown}" if shown else label)
return passed
def head(title: str) -> None:
print(f"\n{title}\n{'-' * len(title)}")
def sha256(path: str) -> str:
h = hashlib.sha256()
with open(path, "rb") as fh:
for chunk in iter(lambda: fh.read(1 << 20), b""):
h.update(chunk)
return h.hexdigest()
def read_csv(path: str) -> "list[dict[str, str]]":
"""Read a CSV into a list of dicts using only the stdlib.
Empty cells come back as the empty string, never as a float NaN. Callers
that care about missing values must test for "" explicitly -- absence must
not be allowed to read as a value.
"""
with open(path, newline="", encoding="utf-8") as fh:
return list(csv.DictReader(fh))
def header(path: str) -> "list[str]":
with open(path, newline="", encoding="utf-8") as fh:
return next(csv.reader(fh), [])
def count_distinct(rows, key_col, group_col):
"""{group value: number of distinct key_col values}."""
acc = {}
for r in rows:
acc.setdefault(r[group_col], set()).add(r[key_col])
return {k: len(v) for k, v in acc.items()}
def count_rows(rows, group_col):
acc = {}
for r in rows:
acc[r[group_col]] = acc.get(r[group_col], 0) + 1
return acc
def main() -> int:
ap = argparse.ArgumentParser()
ap.add_argument("--dir", default=os.path.dirname(os.path.abspath(__file__)))
ap.add_argument("--skip-checksums", action="store_true")
args = ap.parse_args()
root = os.path.abspath(args.dir)
print("=" * 72)
print("Cost-Aware Protocol Routing -- dataset validation")
print("=" * 72)
print(f"Directory: {root}")
# ---- versions -------------------------------------------------------
head("Versions")
mpath = os.path.join(root, "registry/release_manifest.json")
if not os.path.exists(mpath):
ok("registry/release_manifest.json present", False, "missing")
print("\nCannot continue without the manifest.")
return 1
man = json.load(open(mpath))
rv, sv = man.get("dataset_version"), man.get("schema_version")
print(f" release version : {rv}")
print(f" schema version : {sv}")
print(f" generated (UTC) : {man.get('generated_utc')}")
ok("release version matches", rv == EXPECTED_RELEASE_VERSION, f"got {rv}")
ok("schema version matches", sv == EXPECTED_SCHEMA_VERSION, f"got {sv}")
# ---- expected files -------------------------------------------------
head("Expected files")
missing = [f for f in REQUIRED_FILES if not os.path.exists(os.path.join(root, f))]
ok(f"all {len(REQUIRED_FILES)} required files present", not missing,
f"missing: {missing}" if missing else "")
if missing:
print("\nCannot continue with required files missing.")
return 1
# ---- row counts -----------------------------------------------------
head("Row counts")
for rel, exp in EXPECTED_ROWS.items():
p = os.path.join(root, rel)
n = sum(1 for _ in open(p)) if rel.endswith(".jsonl") else sum(1 for _ in open(p)) - 1
ok(f"{rel} == {exp} rows", n == exp, f"got {n}")
m = read_csv(os.path.join(root, "data/matched_labels.csv"))
# ---- coverage -------------------------------------------------------
head("Benchmark coverage")
got_b = count_distinct(m, "problem_id", "benchmark_id")
for b, exp in EXPECTED_BENCHMARKS.items():
ok(f"{b}: {exp} distinct problems", got_b.get(b) == exp, f"got {got_b.get(b)}")
ok("no unexpected benchmark", set(got_b) == set(EXPECTED_BENCHMARKS),
f"unexpected: {sorted(set(got_b) - set(EXPECTED_BENCHMARKS))}")
ok("MaScQA absent (excluded: NonCommercial, not in paper)", "mascqa" not in got_b)
head("Model coverage")
got_m = count_distinct(m, "setting_id", "model")
for mod, exp in EXPECTED_MODELS.items():
ok(f"{mod}: {exp} settings", got_m.get(mod) == exp, f"got {got_m.get(mod)}")
ok("no unexpected solver", set(got_m) == set(EXPECTED_MODELS),
f"unexpected: {sorted(set(got_m) - set(EXPECTED_MODELS))}")
all_settings = {r["setting_id"] for r in m}
ok("Gemma-3-27B scope check absent",
not any("gemma3_27b" in s_ for s_ in all_settings))
head("Setting coverage")
got_s = count_rows(m, "setting_id")
ok("exactly 10 settings", len(got_s) == 10, f"got {len(got_s)}")
for s_, exp in EXPECTED_SETTINGS.items():
ok(f"{s_} == {exp}", got_s.get(s_) == exp, f"got {got_s.get(s_)}")
head("Protocol coverage")
cols = [f"{x}_correct" for x in ["baseline", "single", "per", "broadcast"]]
mcols = header(os.path.join(root, "data/matched_labels.csv"))
ok("all four protocol columns present", all(c in mcols for c in cols),
f"missing: {[c for c in cols if c not in mcols]}")
# Empty string means a missing cell. Absence must not read as a value.
blanks = sum(1 for r in m for c in cols if r[c].strip() == "")
ok("no null protocol outcomes", blanks == 0, f"{blanks} blank cells")
bad = sum(1 for r in m for c in cols if r[c].strip() not in ("0", "1"))
ok("protocol outcomes are strictly 0/1", bad == 0, f"{bad} out of range")
got_lab = {r["oracle_label"] for r in m}
ok("oracle_label uses the documented vocabulary", got_lab <= set(ORACLE_LABELS),
f"unexpected: {sorted(got_lab - set(ORACLE_LABELS))}")
head("Oracle recomputation (Baseline -> Single -> PER -> Broadcast -> None)")
mism = inv = 0
seen = set()
dupes = 0
for r in m:
flags = [r["baseline_correct"], r["single_correct"], r["per_correct"], r["broadcast_correct"]]
rec = "none"
for proto, flag in zip(PROTOCOLS, flags):
if flag == "1":
rec = proto
break
if rec != r["oracle_label"]:
mism += 1
if (r["oracle_label"] == "none") != (r["any_protocol_solved"] == "0"):
inv += 1
k = (r["setting_id"], r["problem_id"])
if k in seen:
dupes += 1
seen.add(k)
ok("oracle recomputation agrees on every row", mism == 0, f"{mism} mismatches")
ok("oracle_label=='none' iff any_protocol_solved==0", inv == 0, f"{inv} violations")
ok("(setting_id, problem_id) unique", dupes == 0, f"{dupes} duplicates")
head("Split integrity")
for rel in ["data/splits/primary_omnimath_splits.csv",
"data/splits/six_setting_splits.csv"]:
rows = read_csv(os.path.join(root, rel))
groups = {}
for r in rows:
g = r.get("setting_id", "(all)")
groups.setdefault(g, {}).setdefault(r["split"], set()).add(r["problem_id"])
bad_n = 0
for _, st in groups.items():
tr, dv, te = st.get("train", set()), st.get("dev", set()), st.get("test", set())
if (tr & dv) or (tr & te) or (dv & te):
bad_n += 1
ok(f"{rel}: train/dev/test pairwise disjoint", bad_n == 0, f"{bad_n} groups overlap")
head("Leakage guard (feature files must carry no outcomes)")
for rel in FEATURE_FILES:
p_ = os.path.join(root, rel)
if rel.endswith(".jsonl"):
with open(p_, encoding="utf-8") as fh:
cs = set(json.loads(fh.readline()).keys())
else:
cs = set(header(p_))
hits = sorted(cs & FORBIDDEN_IN_FEATURES)
ok(f"{rel} carries no outcome/answer column", not hits, f"found {hits}")
head("Confidence probes (two distinct instruments)")
post = read_csv(os.path.join(root, "data/confidence/postanswer_confidence_predictions.csv"))
pre = read_csv(os.path.join(root, "data/confidence/primary_omnimath_confidence_predictions.csv"))
post_types = {r["probe_type"] for r in post}
pre_types = {r["probe_type"] for r in pre}
ok("post-answer file is uniformly probe_type=post_answer_pre_collaboration",
post_types == {"post_answer_pre_collaboration"}, f"got {sorted(post_types)}")
ok("pre-answer file is uniformly probe_type=pre_answer_q1",
pre_types == {"pre_answer_q1"}, f"got {sorted(pre_types)}")
post_eids = [r["example_id"] for r in post]
ok("post-answer example_id is unique", len(set(post_eids)) == len(post_eids),
f"{len(post_eids) - len(set(post_eids))} duplicates")
# Scale and null semantics: a "probability"-named column on a 0-100 scale is
# the kind of thing a reader gets silently wrong, so assert it rather than
# only document it.
post_conf = [float(r["confidence"]) for r in post if r["confidence"].strip() != ""]
post_hi = max(post_conf) if post_conf else 0.0
ok("post-answer `confidence` is on the documented 0-100 scale", post_hi > 1.0,
detail=f"range {min(post_conf)}-{post_hi} over {len(post_conf)} parsed rows",
fail_detail=f"max is {post_hi}: the column looks like a 0-1 probability, "
f"not the documented 0-100 scale")
ok("post-answer `confidence` is blank exactly where parse_ok is false",
sum(1 for r in post if r["confidence"].strip() == "")
== sum(1 for r in post if r["parse_ok"].strip().lower() == "false"))
pre_vals = [float(r["confidence_probability"]) for r in pre
if r["confidence_probability"].strip() != ""]
pre_hi = max(pre_vals) if pre_vals else 0.0
ok("pre-answer `confidence_probability` is on the documented 0-100 scale",
pre_hi > 1.0,
detail=f"range {min(pre_vals)}-{pre_hi} over {len(pre_vals)} parsed rows",
fail_detail=f"max is {pre_hi}: the column looks like a 0-1 probability, "
f"not the documented 0-100 scale")
fb = [r for r in pre if r["used_fallback"].strip().lower() == "true"]
fb_with_value = sum(1 for r in fb if r["confidence_probability"].strip() != "")
ok("pre-answer fallback rows carry NO confidence value (documented as null)",
fb_with_value == 0, f"{fb_with_value} of {len(fb)} fallback rows have a value")
print(f" {len(fb)} fallback rows are null; treat a null as ESCALATE "
f"(see docs/schema.md -- other readings give 73.76% or 60.76% "
f"against the published 78.0%)")
xw = read_csv(os.path.join(root, "registry/example_id_crosswalk.csv"))
xw_keys = {(r["setting_id"], r["example_id"], r["problem_id"]) for r in xw}
resolved = sum(1 for r in post
if (r["setting_id"], r["example_id"], r["problem_id"]) in xw_keys)
ok("crosswalk resolves every post-answer example_id", resolved == len(post),
f"{resolved}/{len(post)}")
matched_keys = {(r["setting_id"], r["problem_id"]) for r in m}
joined = sum(1 for r in post
if (r["setting_id"], r["example_id"], r["problem_id"]) in xw_keys
and (r["setting_id"], r["problem_id"]) in matched_keys)
ok("post-answer predictions join to matched outcomes via the crosswalk",
joined == len(post), f"{joined}/{len(post)}")
head("Cost layer (one setting only)")
costs = read_csv(os.path.join(root, "data/costs/omnimath_per_protocol_costs.csv"))
cset = {r["setting_id"] for r in costs}
ok("costs cover exactly one setting",
cset == {"omnimath2__competition_math_4181__gpt_oss_120b"}, f"got {sorted(cset)}")
cids = [r["problem_id"] for r in costs]
ok("cost problem_id is unique", len(set(cids)) == len(cids),
f"{len(cids) - len(set(cids))} duplicates")
om = {r["problem_id"] for r in m
if r["setting_id"] == "omnimath2__competition_math_4181__gpt_oss_120b"}
ok("every cost row resolves to a released problem", set(cids) <= om,
f"{len(set(cids) - om)} unresolved")
ok("26 of 4181 problems omitted (duplicate-text groups; see TODO.md)",
len(om - set(cids)) == 26, f"{len(om - set(cids))} omitted")
tokcols = [f"{x}_total_tokens" for x in ["baseline", "single", "per", "broadcast"]]
blank = sum(1 for r in costs for c in tokcols if r[c].strip() == "")
ok("no blank token totals", blank == 0, f"{blank} blanks")
mean_base = sum(int(r["baseline_total_tokens"]) for r in costs) / len(costs)
print(f" mean Baseline tokens over the {len(costs)} emitted rows: {mean_base:,.1f}")
print(" (the paper's figure over all 4,181 problems is 18,385.4 -- the gap is the")
print(" 26 omitted rows, not a disagreement. See docs/schema.md.)")
ok("mean Baseline tokens is the documented 18,432.0",
abs(mean_base - 18432.0) < 0.05, f"got {mean_base:.1f}")
# ---- checksums ------------------------------------------------------
head("Checksums")
cpath = os.path.join(root, "registry/checksums.sha256")
expected = {}
for line in open(cpath):
line = line.strip()
if not line or line.startswith("#"):
continue
digest, rel = line.split(None, 1)
expected[rel.strip()] = digest
print(f" {len(expected)} entries in registry/checksums.sha256")
if args.skip_checksums:
print(" (skipped by --skip-checksums)")
else:
bad, absent = [], []
for rel, digest in sorted(expected.items()):
p = os.path.join(root, rel)
if not os.path.exists(p):
absent.append(rel)
elif sha256(p) != digest:
bad.append(rel)
ok("every checksummed file is present", not absent, f"missing: {absent}")
ok("every checksum matches", not bad, f"mismatched: {bad}")
on_disk = set()
skipped_infra = 0
for dp, _, fs in os.walk(root):
for fn in fs:
r = os.path.relpath(os.path.join(dp, fn), root)
if r == "registry/checksums.sha256":
continue
if is_infrastructure(r):
skipped_infra += 1
continue
on_disk.add(r)
if skipped_infra:
print(f" ({skipped_infra} tool-created file(s) ignored: download cache, "
f"version-control and OS metadata)")
extra = sorted(on_disk - set(expected))
if extra:
# An unexpected extra file is not corruption, so this stays a warning.
WARNS.append(f"{len(extra)} file(s) present but not checksummed: "
f"{abbreviate(extra)}")
print(f" [WARN] {len(extra)} file(s) not in checksums.sha256: "
f"{abbreviate(extra)}")
# ---- optional artifacts --------------------------------------------
head("Optional artifacts (absent by design)")
for rel, why in OPTIONAL_ARTIFACTS.items():
present = os.path.exists(os.path.join(root, rel.rstrip("/")))
print(f" [{'present' if present else 'ABSENT '}] {rel} -- {why}")
print("\n NOTE: Baseline final-answer text is NOT in this release. The post-answer")
print(" confidence probe consumed it, so the probe CANNOT be re-run from this")
print(" release alone. data/probe_inputs.jsonl is an identifier manifest, not a")
print(" runnable prompt set. See docs/reconstruction.md and TODO.md.")
# ---- result ---------------------------------------------------------
print("\n" + "=" * 72)
if FAILS:
print(f"RESULT: FAILED -- {len(FAILS)} check(s) did not pass")
for f in FAILS:
print(f" - {f}")
print("=" * 72)
return 1
print("RESULT: PASSED -- all checks succeeded")
if WARNS:
print(f"({len(WARNS)} warning(s))")
for w in WARNS:
print(f" ! {w}")
print("=" * 72)
return 0
if __name__ == "__main__":
sys.exit(main())
|