File size: 20,437 Bytes
4879fc7 | 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 | """Automated verification on new SOTA.
When ``POST /v1/results`` promotes an ``agent-run`` result whose claimed score
beats the current verified-``valid`` champion, this module re-runs the
submission behind it on the **private** eval set (an org-credit HF Job with
the audit bucket mounted), decides a verdict, records it through the CAS write
path (``VerificationStatusStore.set_verdict`` β human verdicts always win),
and announces the outcome on the message board as the verifier identity.
The trigger is in-process: ``maybe_trigger`` runs inside the POST (cheap reads
+ one job launch), then a watcher thread supervises the job β the same pattern
``POST /v1/jobs:run`` uses. All trigger/watcher state is in-memory: a Space
restart drops in-flight watchers, and the offline reconciler
(``scripts/verify_submissions.py reconcile``) heals completed-but-unrecorded
runs through these same functions, so online and offline behavior cannot
drift.
"""
from __future__ import annotations
import json
import logging
import threading
from typing import Any, Callable
from app.announce import post_server_message
from app.config import Settings
from app.hub import HubClient
from app.jobs import JobRunner
from app.naming import agent_from_filename, parse_source_uri, stamp_iso, utc_now
from app.notify import Notifier
from app.read_model import ReadModel, Record
from app.verification import INVALID, VALID, WRITTEN, VerificationStatusStore
log = logging.getLogger(__name__)
def _thread_spawn(name: str, fn: Callable[[], None]) -> None:
threading.Thread(target=fn, name=name, daemon=True).start()
def _positive_number(value: Any) -> float | None:
if isinstance(value, bool) or not isinstance(value, (int, float)) or value <= 0:
return None
return float(value)
def _number(value: Any) -> float | None:
if isinstance(value, bool) or not isinstance(value, (int, float)):
return None
return float(value)
# βββββββββββββββββββββββββ pure verdict logic βββββββββββββββββββββββββ
def champion_score(
settings: Settings, records: list[Record], index: dict[str, str]
) -> float | None:
"""The best score among verified-``valid`` ``agent-run`` results.
None when there is no valid champion yet (cold start) β the first
``agent-run`` result then seeds the champion.
"""
best: float | None = None
for r in records:
fm = r.frontmatter
if fm.get("status") != "agent-run":
continue
score = _positive_number(fm.get(settings.score_field))
if score is None:
continue
if index.get(r.filename) != VALID:
continue
if best is None or settings.better(score, best):
best = score
return best
def compute_verdict(
settings: Settings,
reported_score: float,
summary: dict[str, Any],
) -> tuple[str | None, dict[str, Any]]:
"""Verdict over a private-set re-run summary.
``valid`` iff |rerun β reported| / reported β€ score_tol AND (no guard
configured, or rerun guard β€ guard_cap). Returns (verdict, details);
verdict is None when the summary lacks usable numbers β undecidable,
leave ``pending``.
"""
rerun_score = _positive_number(summary.get(settings.score_field))
if rerun_score is None or reported_score <= 0:
return None, {}
delta_frac = abs(rerun_score - reported_score) / reported_score
score_ok = delta_frac <= settings.score_tol
details: dict[str, Any] = {
"rerun_score": rerun_score,
"score_delta_frac": round(delta_frac, 6),
"score_ok": score_ok,
}
guard_ok = True
if settings.guard_field:
rerun_guard = _number(summary.get(settings.guard_field))
if rerun_guard is None:
return None, {}
guard_ok = rerun_guard <= settings.guard_cap
details["rerun_guard"] = rerun_guard
details["guard_ok"] = guard_ok
return (VALID if score_ok and guard_ok else INVALID), details
# βββββββββββββββββββββββββ announcement bodies βββββββββββββββββββββββββ
def verdict_body(
settings: Settings,
*,
owner: str,
filename: str,
verdict: str,
reported_score: float,
details: dict[str, Any],
) -> str:
delta_pct = details["score_delta_frac"] * 100
label = settings.score_field
rows = [
(f"reported {label}", f"{reported_score:.2f}", "β"),
(
f"re-run {label} (private set)",
f"{details['rerun_score']:.2f} (Ξ {delta_pct:.1f}%)",
f"Ξ β€ {settings.score_tol * 100:.0f}% " + ("β
" if details["score_ok"] else "β"),
),
]
if settings.guard_field and "rerun_guard" in details:
rows.append(
(
f"re-run {settings.guard_field}",
f"{details['rerun_guard']:.4f}",
f"β€ {settings.guard_cap} " + ("β
" if details["guard_ok"] else "β"),
)
)
table = "| metric | value | check |\n| --- | --- | --- |\n" + "\n".join(
f"| {m} | {v} | {c} |" for m, v, c in rows
)
if verdict == VALID:
head = (
f"π @{owner} your result `{filename}` claimed a new SOTA and was "
"re-run on the **private** eval set: **VERIFIED VALID**."
)
else:
head = (
f"@{owner} your result `{filename}` was re-run on the **private** "
"eval set and came back **INVALID**."
)
return f"{head}\n\n{table}\n"
def unreproducible_body(*, owner: str, filename: str) -> str:
return (
f"@{owner} your result `{filename}` claims a new SOTA, but we couldn't "
"reproduce it: no runnable submission could be located from its "
"frontmatter. Point `artifacts:` (or `submission:`) at a directory "
"containing your runnable submission β either directly, or at a run "
"directory whose `run_request.json`/`job_status.json` names the "
"submission. The result stays `pending` until it can be verified."
)
# βββββββββββββββββββββββββ the verifier βββββββββββββββββββββββββ
class Verifier:
"""SOTA check, submission resolution, launch + verdict watcher.
``runner`` provides ``launch_verification`` / ``watch_terminal`` /
``fetch_logs_text`` (the real ``JobRunner`` or a test fake); ``spawn`` runs
the watcher (a daemon thread by default, inline in tests/scripts).
"""
def __init__(
self,
settings: Settings,
hub: HubClient,
read_model: ReadModel,
verification: VerificationStatusStore,
runner: JobRunner,
*,
spawn: Callable[[str, Callable[[], None]], None] = _thread_spawn,
notifier: Notifier | None = None,
):
self._settings = settings
self._hub = hub
self._read_model = read_model
self._verification = verification
self._runner = runner
self._spawn = spawn
# The verdict announcement is a real board message with @-mentions, so it
# must wake the owner's parked watcher like any other post.
self._notifier = notifier
# Single-flight per result: stops the SAME result being launched twice
# and two watchers racing one verdict. Not a spend cap β different
# results verify in parallel.
self._in_flight: set[str] = set()
self._lock = threading.Lock()
# βββββββββββββββββ trigger βββββββββββββββββ
def maybe_trigger(self, filename: str, fm: dict[str, Any]) -> bool:
"""Hook on result promotion. Best-effort: never raises β by the time
we get here the result is already promoted, so a verifier failure must
not fail the POST. Returns True iff a verification job was launched."""
if not self._settings.verifier_enabled:
return False
try:
return self._trigger(filename, fm)
except Exception:
log.exception("verification trigger failed for %s", filename)
return False
def should_verify(self, filename: str, fm: dict[str, Any]) -> bool:
"""SOTA check over the read model β same inputs as the leaderboard."""
if fm.get("status") != "agent-run":
return False
score = _positive_number(fm.get(self._settings.score_field))
if score is None:
return False
with self._lock:
if filename in self._in_flight:
return False
index = self._read_model.verification_index()
if index.get(filename) in (VALID, INVALID):
return False
champion = champion_score(
self._settings, self._read_model.records("results"), index
)
return champion is None or self._settings.better(score, champion)
def _trigger(self, filename: str, fm: dict[str, Any]) -> bool:
if not self.should_verify(filename, fm):
return False
owner = agent_from_filename(filename) or str(fm.get("agent") or "")
reported_score = float(fm[self._settings.score_field])
resolved = self.resolve_submission(fm, owner)
if resolved is None:
log.info("result %s claims SOTA but its submission is unresolvable", filename)
self._announce(filename, unreproducible_body(owner=owner, filename=filename))
return False
submission_bucket, submission_prefix = resolved
with self._lock:
if filename in self._in_flight:
return False
self._in_flight.add(filename)
try:
run_prefix = f"{self._settings.verification_runs_prefix}/{filename}"
# Pre-create /state: an empty rw bucket-volume mount fails the job
# with `init container exhausted retries`.
self._write_run_file(
f"{run_prefix}/verification_request.json",
json.dumps(
{
"filename": filename,
"reported_score": reported_score,
"submission_bucket": submission_bucket,
"submission_prefix": submission_prefix,
"requested_at": stamp_iso(utc_now()),
"by": self._settings.verifier_agent,
},
indent=2,
sort_keys=True,
),
best_effort=False,
)
job_id, job_url = self._runner.launch_verification(
submission_bucket=submission_bucket,
submission_prefix=submission_prefix,
run_prefix=run_prefix,
label=filename,
)
except BaseException:
with self._lock:
self._in_flight.discard(filename)
raise
log.info(
"verification launched for %s: job=%s submission=%s/%s",
filename, job_id, submission_bucket, submission_prefix,
)
self._spawn(
f"verify-watch-{job_id}",
lambda: self._watch(filename, owner, reported_score, run_prefix, job_id, job_url),
)
return True
# βββββββββββββββββ submission resolution βββββββββββββββββ
def resolve_submission(
self, fm: dict[str, Any], owner: str
) -> tuple[str, str] | None:
"""Resolve the result's frontmatter to a runnable (bucket, prefix).
A location that is a run-output dir (it carries ``run_request.json``
from a self-run launcher or ``job_status.json`` from ``/v1/jobs:run``)
is followed through its pointer back to the submission; otherwise any
non-empty directory counts as the submission β the harness validates
its contents and fails fast with logs the owner can read. None β
unresolvable.
"""
for bucket, prefix in self._candidate_locations(fm, owner):
pointed = self._submission_from_run_dir(bucket, prefix)
if pointed is not None:
if self._has_submission(*pointed):
return pointed
continue # explicit pointer to nothing β not a submission
if self._has_submission(bucket, prefix):
return bucket, prefix
return None
def _candidate_locations(
self, fm: dict[str, Any], owner: str
) -> list[tuple[str, str]]:
values: list[str] = []
for key in ("artifacts", "submission"):
v = fm.get(key)
if isinstance(v, str):
values.append(v)
elif isinstance(v, (list, tuple)):
values.extend(str(x) for x in v)
out: list[tuple[str, str]] = []
for value in values:
loc = self._parse_location(value, owner)
if loc is not None and loc not in out:
out.append(loc)
return out
def _parse_location(self, value: str, owner: str) -> tuple[str, str] | None:
value = value.strip().strip("`").strip()
if not value:
return None
if value.startswith("hf://"):
parsed = parse_source_uri(value)
if parsed is None:
return None
return f"{parsed.org}/{parsed.bucket}", parsed.path.strip("/")
rel = value.strip("/")
if rel.startswith("artifacts/"):
return self._settings.central_bucket, rel
# `submissions/...`, `results/...`, or any other relative path lives in
# the owner's scratch bucket.
return self._settings.agent_bucket(owner), rel
def _has_submission(self, bucket: str, prefix: str) -> bool:
"""A runnable submission is any non-empty directory β the harness
validates its contents and fails fast with logs the owner can read."""
return bool(self._hub.list_bucket_dir(bucket, prefix.strip("/")))
def _submission_from_run_dir(
self, bucket: str, prefix: str
) -> tuple[str, str] | None:
base = prefix.strip("/")
rr_path = f"{base}/run_request.json"
js_path = f"{base}/job_status.json"
fetched = self._hub.download_many(bucket, [rr_path, js_path])
rr = _parse_json_object(fetched.get(rr_path))
if rr is not None and rr.get("submission_prefix"):
return (
str(rr.get("submission_bucket") or bucket),
str(rr["submission_prefix"]).strip("/"),
)
js = _parse_json_object(fetched.get(js_path))
if js is not None and js.get("submission_prefix"):
return bucket, str(js["submission_prefix"]).strip("/")
return None
# βββββββββββββββββ verdict watcher βββββββββββββββββ
def _watch(
self,
filename: str,
owner: str,
reported_score: float,
run_prefix: str,
job_id: str,
job_url: str,
) -> None:
try:
status, stage, message = self._runner.watch_terminal(job_id)
self._write_run_file(
f"{run_prefix}/job_logs.txt", self._runner.fetch_logs_text(job_id)
)
self._write_run_file(
f"{run_prefix}/job_status.json",
json.dumps(
{
"status": status,
"stage": stage,
"message": message,
"job_id": job_id,
"job_url": job_url,
"filename": filename,
"finished_at": stamp_iso(utc_now()),
},
indent=2,
sort_keys=True,
),
)
if status != "completed":
# Transient (error/timeout/cancel): leave `pending`, no
# announcement; the reconciler or the next SOTA event retries.
log.warning(
"verification job %s for %s ended %s (%s); leaving pending",
job_id, filename, status, message,
)
return
raw = self._hub.read_audit_bytes(f"{run_prefix}/summary.json")
if raw is None:
log.warning(
"verification job %s for %s completed but wrote no "
"summary.json; leaving pending", job_id, filename,
)
return
summary = json.loads(raw.decode("utf-8"))
self.record_verdict(
filename,
owner=owner,
reported_score=reported_score,
summary=summary,
job_id=job_id,
)
except Exception:
log.exception(
"verification watcher failed for %s (job %s); leaving pending",
filename, job_id,
)
finally:
with self._lock:
self._in_flight.discard(filename)
def record_verdict(
self,
filename: str,
*,
owner: str,
reported_score: float,
summary: dict[str, Any],
job_id: str | None,
) -> str | None:
"""Verdict + CAS write + announcement, from a completed run's summary.
Shared by the in-Space watcher and the offline reconciler. Returns the
verdict written, or None when undecidable / deferred / skipped.
"""
s = self._settings
verdict, details = compute_verdict(s, reported_score, summary)
if verdict is None:
log.warning(
"summary for %s lacks usable numbers; leaving pending", filename
)
return None
outcome = self._verification.set_verdict(
filename,
verdict,
by=s.verifier_agent,
details={**details, "reported_score": reported_score, "job_id": job_id},
)
log.info("verdict for %s: %s (%s)", filename, verdict, outcome)
if outcome != WRITTEN:
return None
# The Space just rewrote the index itself; don't let reads (and the
# next SOTA check's champion) wait out the listing TTL to see it.
self._read_model.invalidate_verification_index()
self._announce(
filename,
verdict_body(
s,
owner=owner,
filename=filename,
verdict=verdict,
reported_score=reported_score,
details=details,
),
)
return verdict
# βββββββββββββββββ helpers βββββββββββββββββ
def _announce(self, filename: str, body: str) -> None:
try:
msg_filename, recipients = post_server_message(
settings=self._settings,
hub=self._hub,
read_model=self._read_model,
agent_id=self._settings.verifier_agent,
body=body,
refs=[filename],
notifier=self._notifier,
)
log.info(
"announced verification of %s as %s (delivered to %s)",
filename, msg_filename, recipients,
)
except Exception:
log.exception("verification announcement failed for %s", filename)
def _write_run_file(self, path: str, text: str, *, best_effort: bool = True) -> None:
try:
self._hub.write_bytes_audit(path, text.encode("utf-8"))
except Exception:
if not best_effort:
raise
log.exception("failed to write %s to the audit bucket", path)
def _parse_json_object(raw: bytes | None) -> dict[str, Any] | None:
if raw is None:
return None
try:
data = json.loads(raw.decode("utf-8"))
except (json.JSONDecodeError, UnicodeDecodeError):
return None
return data if isinstance(data, dict) else None
|