Spaces:
Sleeping
Sleeping
File size: 17,088 Bytes
e8ef5c3 | 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 | # Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
"""
Long-policy task bank for Prompt Golf.
These tasks exist to stress-test prompt compression on a real-world
pattern: an organization has a multi-page policy (ad standards, content
moderation, financial disclosures) and needs an LLM-based classifier
that decides whether a piece of content complies. The naive prompt is
the *entire policy verbatim* — easily 1000-2000 tokens. The trained
prompt golf agent must compress that into a tight classifier prompt
that preserves the load-bearing rules.
Why these tasks are valuable:
- Long policy text = aggressive compression target (10×+ ratio).
- Multi-class decisions test that the agent identifies the
*hierarchy* of rules, not just keywords.
- Real businesses pay for policy compression today — every byte of
system prompt costs $$ at scale.
Each task ships with:
- description: the full multi-clause policy (1000+ tokens)
- train_examples: 3 (content_description, decision) pairs
- test_examples: 6 hidden (content_description, decision) pairs
- scorer: exact_label (decision is from a closed vocabulary)
- budget_tokens: 250 (aggressive — verbose policy is ~1200 tokens)
"""
from __future__ import annotations
try:
from .tasks import TaskSpec
except ImportError:
from server.tasks import TaskSpec
TASKS_POLICY: dict[str, TaskSpec] = {}
def _add(task: TaskSpec) -> None:
TASKS_POLICY[task.task_id] = task
def list_task_ids_policy() -> list[str]:
return list(TASKS_POLICY.keys())
# ============================================================================
# 1. Ad creative policy compliance (MSN/Bing-style)
# ============================================================================
_MSN_AD_POLICY = """\
MSN AD CREATIVE POLICY (effective 2025)
This document describes the standards every ad creative must meet \
before serving on the MSN advertising network. Each submitted creative \
must be classified into exactly one of three buckets: ALLOW (compliant, \
serve as-is), DISALLOW (violates a hard prohibition, do not serve), or \
REVIEW (potentially compliant but requires manual editorial review or \
the addition of mandatory disclosures before serving).
SECTION A — HARD PROHIBITIONS (always DISALLOW)
A.1 Illegal goods or services. Any creative promoting goods or \
services that are illegal under United States federal law or in the \
country where the creative will serve is prohibited. This includes \
unlicensed pharmaceuticals, controlled substances, illegal weapons, \
counterfeit goods, and pirated content.
A.2 Tobacco, e-cigarettes, vaping products. Promotion of any tobacco \
or nicotine-delivery product is prohibited in the United States, \
European Union, Australia, and Singapore. Cessation aids prescribed by \
licensed clinicians are exempt.
A.3 Adult content. Sexually explicit material, escort services, and \
adult entertainment are prohibited.
A.4 Misleading medical claims. Any creative that claims to cure, \
prevent, or guarantee treatment outcomes for a serious medical \
condition (cancer, diabetes, HIV/AIDS) is prohibited unless backed by \
FDA approval cited in the creative.
A.5 Cryptocurrency speculation. Promotion of unregistered initial coin \
offerings, leveraged crypto trading without risk disclosure, or any \
'guaranteed returns' framing in the digital-asset category.
SECTION B — RESTRICTED CONTENT (REVIEW unless conditions met)
B.1 Alcohol. Permitted only when (a) creative includes 'Drink \
Responsibly' or jurisdiction-equivalent disclaimer, (b) campaign is \
restricted to ages 21+ in the US (18+ in countries with lower legal \
drinking age), and (c) target country permits alcohol advertising. If \
all three conditions are explicitly stated, ALLOW; otherwise REVIEW.
B.2 Gambling and games of chance. Permitted only with (a) state or \
country-level operator license number visible in the creative, and (b) \
problem-gambling helpline displayed. Both required for ALLOW.
B.3 Financial services and investment products. Stock trading, \
brokerage, and investment-management ads require an explicit risk \
disclosure clause ('investments may lose value' or equivalent). Loans \
and credit cards require APR ranges. If disclosure is present, ALLOW; \
otherwise REVIEW for editorial sign-off.
B.4 Political and issue advertising. Any creative referencing a \
political candidate, ballot measure, or contested social issue must \
have advertiser identity verified through MSN's political verification \
program. Without verification: REVIEW.
B.5 Health and wellness (non-medical). Supplements, fitness programs, \
and weight-loss products are permitted only when claims are factual and \
do not promise specific outcomes ('lose 20 lbs guaranteed' is \
prohibited; 'supports healthy weight management' is permitted).
SECTION C — FORMAT AND PRESENTATION STANDARDS
C.1 Creative must clearly identify the advertiser. Anonymous or \
deceptively branded creatives are REVIEW.
C.2 No fake-system-UI elements. Creatives that imitate browser \
warnings, OS notifications, or user-interface chrome to induce clicks \
are DISALLOW.
C.3 No autoplay audio with sound on by default.
C.4 All claims of superiority ('#1', 'best in class') must be \
independently verifiable; otherwise REVIEW.
DECISION FORMAT
For each input creative description, output exactly one label, \
lowercase, no punctuation, no explanation: 'allow', 'disallow', or \
'review'. The label must be on the first line of the output and \
nothing else.\
"""
_add(TaskSpec(
task_id="policy_msn_ad_creative",
category="policy_compression",
description=_MSN_AD_POLICY,
scorer="exact_label",
train_examples=[
("Banner ad for Heineken beer, copy includes 'Drink Responsibly', "
"campaign restricted to ages 21+, serving in the United States.",
"allow"),
("Banner ad promoting an unregistered ICO promising 'guaranteed "
"300% returns in 90 days'.",
"disallow"),
("Banner ad for Robinhood stock trading app, copy mentions "
"'commission-free trading' but no risk disclosure visible.",
"review"),
],
test_examples=[
("Banner ad for a vape pen brand, campaign serving in the US.",
"disallow"),
("Native ad for an online casino, includes operator license "
"number 'NV-2189' and 1-800-GAMBLER helpline.",
"allow"),
("Display ad for Nike running shoes, no special claims.",
"allow"),
("Sponsored post for a weight-loss supplement, copy says 'lose 20 "
"pounds in 30 days, guaranteed!'.",
"disallow"),
("Banner ad endorsing Senator Park's re-election campaign, no "
"advertiser-verification badge present.",
"review"),
("Display ad for Wells Fargo personal loans, copy mentions APR "
"range '6.99%-19.99%' clearly.",
"allow"),
],
budget_tokens=250,
difficulty="hard",
tags=["policy", "compression", "ad-tech", "long-context"],
))
# ============================================================================
# 2. Content moderation policy (social-platform style)
# ============================================================================
_CONTENT_MOD_POLICY = """\
COMMUNITY CONTENT POLICY (v3.2)
Every user-submitted post on this platform is reviewed against the \
following rules. Each post is classified into exactly one of four \
buckets: KEEP (post complies, do not act), WARN (post is borderline; \
attach a sensitive-content interstitial but leave it visible), REMOVE \
(post violates a hard rule; delete it and notify the author), or \
ESCALATE (post may involve real-world harm or legal risk; route to a \
human reviewer immediately).
PART 1 — HARD REMOVAL CRITERIA (always REMOVE)
1.1 Doxxing. Posts that publish another user's private contact \
information (home address, personal phone, employer in a hostile \
context) without that user's consent.
1.2 CSAM and child exploitation. Any sexualized content involving \
minors, or content that grooms, recruits, or trafficks minors. These \
posts also trigger ESCALATE simultaneously.
1.3 Direct credible threats of violence. Posts that name a specific \
person, place, or scheduled event AND describe an intended violent act.
1.4 Coordinated harassment. Posts directing multiple users to attack a \
named target (pile-on, brigade, raid).
1.5 Illegal goods marketplace. Selling, buying, or arranging the \
exchange of regulated firearms without a license, controlled substances \
without a prescription, or services rendered for criminal purposes.
PART 2 — ESCALATE TO HUMAN REVIEWER (always ESCALATE)
2.1 First-person suicidal ideation or imminent self-harm. Route to \
crisis-resource workflow within 15 minutes.
2.2 Reports of ongoing real-world emergencies (active shooter, \
kidnapping in progress, hostage situation) — route to law enforcement \
liaison.
2.3 Coordinated influence operations or election-interference patterns \
flagged by the platform-integrity model.
2.4 Legal compliance gray zones (DMCA, GDPR right-to-erasure, court \
preservation orders).
PART 3 — WARN (interstitial but keep)
3.1 Graphic but newsworthy content. Photojournalism from conflict \
zones, accident scenes, or medical procedures shown for educational \
purposes. Apply sensitive-content overlay; allow click-through.
3.2 Suicide / self-harm references in the third person, news, fiction, \
or recovery contexts (NOT first-person ideation — that's ESCALATE).
3.3 Extremely strong language directed at a public figure (politicians, \
celebrities) but stopping short of credible threat or doxxing.
3.4 Adult nudity in clearly artistic, educational, or medical contexts \
(NOT pornographic).
PART 4 — KEEP DEFAULTS
Posts that don't trigger Part 1, 2, or 3 are KEEP. This includes \
ordinary disagreement, strong political opinion, profanity not directed \
at a specific user, satire and parody, and historical or scientific \
discussion of sensitive topics.
OUTPUT FORMAT
For each post description, output exactly one label, lowercase, no \
punctuation, no explanation: 'keep', 'warn', 'remove', or 'escalate'. \
First line only.\
"""
_add(TaskSpec(
task_id="policy_content_moderation",
category="policy_compression",
description=_CONTENT_MOD_POLICY,
scorer="exact_label",
train_examples=[
("Post that names a specific journalist by full name and posts "
"their home address, calling for 'accountability'.",
"remove"),
("News photograph from a conflict zone showing wounded civilians, "
"captioned with factual reporting context.",
"warn"),
("Post containing strong profanity directed at a politician's "
"policy stance, no doxxing or threats.",
"keep"),
],
test_examples=[
("First-person post saying 'I have a plan and pills, tonight is "
"the night' with no recovery framing.",
"escalate"),
("Satirical post mocking a CEO's recent quarterly earnings call, "
"uses sarcasm but no slurs or threats.",
"keep"),
("Post offering to sell an unregistered AR-15 to whoever sends "
"the highest bid via DM.",
"remove"),
("Documentary-style post about historical lynchings in the 20th "
"century US, includes one period photograph.",
"warn"),
("Post coordinating dozens of accounts to flood a single user's "
"mentions with insults over a public dispute.",
"remove"),
("Post sharing a recipe for homemade lemon cake with no other "
"context.",
"keep"),
],
budget_tokens=250,
difficulty="hard",
tags=["policy", "compression", "moderation", "long-context"],
))
# ============================================================================
# 3. Financial-disclosure compliance (broker/dealer style)
# ============================================================================
_FINREG_POLICY = """\
FINRA-STYLE COMMUNICATION REVIEW POLICY (excerpted, simplified for \
classifier training)
A registered representative or firm-affiliated person at a broker-dealer \
must have all written communications with retail customers reviewed \
before they are sent. Each communication is classified into one of four \
buckets: APPROVED (compliant, send), HOLD (return to author for \
correction of a fixable issue), REJECT (substantive violation, do not \
send under any revision), or ESCALATE (requires compliance officer \
sign-off because of complexity, novel product, or potential customer \
harm).
PART I — REJECT CATEGORIES (always REJECT)
1.1 Performance guarantees. Any statement that the customer's \
investment 'cannot lose' or will 'guarantee' a specific return over a \
specific time horizon. Phrases like 'risk-free' applied to anything \
other than US Treasuries are REJECT.
1.2 Selective performance highlighting (cherry-picking). Citing only \
the best historical years of a fund without stating the comparable \
benchmark return AND the worst year in the same period.
1.3 Promissory language about future appreciation. 'Will' double, \
'must' rise, 'has to' recover.
1.4 Unsuitable product recommendations to vulnerable customers. \
Recommending leveraged ETFs, options strategies, or crypto futures to \
customers under 25 or over 75, or with stated 'capital preservation' \
objectives, in a public-facing communication.
PART II — HOLD (revisable)
2.1 Missing standard disclosures. Mutual fund pitches missing the \
'past performance does not guarantee future results' line, or annuity \
pitches missing the surrender-charge schedule reference.
2.2 Missing FINRA Rule 2210 footer (firm name, registration status, \
and contact info).
2.3 Fee disclosures present but not 'clear and prominent' (font size \
or contrast inadequate).
PART III — ESCALATE (compliance officer required)
3.1 New-issue securities communications.
3.2 Communications mentioning private placements, structured products, \
or non-traded REITs.
3.3 References to options strategies more complex than basic covered \
calls.
3.4 Communications targeting accounts opened in the past 90 days \
(new-customer suitability review).
PART IV — APPROVED DEFAULTS
Communications that include all required disclosures, do not promise \
returns, do not cherry-pick performance data, and concern only \
plain-vanilla products (broad-market index funds, money market funds, \
US Treasuries) for customers in the firm's standard suitability bands. \
APPROVED.
OUTPUT FORMAT
Output exactly one lowercase label, no punctuation, no explanation: \
'approved', 'hold', 'reject', or 'escalate'. First line only.\
"""
_add(TaskSpec(
task_id="policy_finreg_communication_review",
category="policy_compression",
description=_FINREG_POLICY,
scorer="exact_label",
train_examples=[
("Email pitch promising clients that the firm's growth fund "
"'will deliver double-digit returns over the next decade'.",
"reject"),
("Brochure for an S&P 500 index fund, includes the standard "
"'past performance does not guarantee future results' "
"disclosure and FINRA footer.",
"approved"),
("Newsletter about a non-traded REIT being added to the firm's "
"platform, includes no obvious violations but is a new "
"structured-product offering.",
"escalate"),
],
test_examples=[
("Mailer for a money-market fund, omits the 'past performance' "
"line but otherwise compliant; FINRA footer present.",
"hold"),
("Email to a 78-year-old retiree with stated 'capital "
"preservation' goal, recommending a 3x leveraged tech ETF.",
"reject"),
("Email pitching a new private placement to a sophisticated "
"investor, with all standard risk disclosures present.",
"escalate"),
("Quarterly note about US Treasury yields, no recommendations, "
"all firm disclosures present.",
"approved"),
("Brochure highlighting only the fund's three best years out of "
"ten, omitting the worst year and benchmark return.",
"reject"),
("Email pitching a covered-call strategy to a customer who "
"opened the account 30 days ago.",
"escalate"),
],
budget_tokens=250,
difficulty="hard",
tags=["policy", "compression", "finance", "long-context"],
))
if __name__ == "__main__":
from collections import Counter
print(f"tasks_policy: {len(TASKS_POLICY)} tasks")
for tid, spec in TASKS_POLICY.items():
desc_words = len(spec.description.split())
print(f" {tid:42s} category={spec.category:22s} "
f"budget={spec.budget_tokens} desc_words≈{desc_words}")
|