Spaces:
Sleeping
Sleeping
File size: 19,564 Bytes
a6f0611 | 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 | """
Round 2 shared evaluation runner.
Deterministic loop + two LLM decision points per suspicious account.
Framework logic (deterministic, no LLM):
- GET_POLICY as step 0
- Loop over visible accounts sorted by risk_score desc
- INSPECT if not yet inspected
- INVESTIGATE_NETWORK on very high-risk suspects (risk >= 0.80, once)
- SUBMIT at end
Per-account LLM decision points:
- DP1 (tool selection): "reverse_image_search" | "analyze_bio" | "check_ip" | "done"
- DP2 (flag decision): "flag" | "skip"
Entry point:
run_evaluation(model_name, call_llm, platform, base_url, tasks, seeds, out_path)
"""
from __future__ import annotations
import json
import os
import sys
import time
from dataclasses import dataclass, asdict, field
from pathlib import Path
from typing import Callable, Dict, List, Optional
# Make parent (fake_gang_env/) importable so the local client + models resolve
# AHEAD of any same-named modules in site-packages. Always insert at index 0
# (and drop any pre-cached aliases) — a no-op `if not in sys.path` check is
# unsafe because the path may already be present at lower priority.
_PARENT = Path(__file__).resolve().parent.parent
sys.path.insert(0, str(_PARENT))
for _stale in ("models", "client"):
sys.modules.pop(_stale, None)
from client import FakeGangEnvClient, StepResult # noqa: E402
from models import ActionType, FakeGangAction # noqa: E402
# ---------------------------------------------------------------------------
# Prompts for the two decision points
# ---------------------------------------------------------------------------
DP1_PROMPT = """You are evaluating a suspicious account. Pick the next tool to call, or 'done' if evidence is sufficient.
PLATFORM: {platform} | primary signal: {primary_signal} | threshold θ*: {threshold:.3f}
ACCOUNT: {account_id} | risk_score: {risk:.3f} | hub_legitimacy: {hub:.2f}
Revealed signals (None = not yet revealed):
photo_reuse_score: {photo}
bio_template_score: {bio}
ip_cluster_signal: {ip}
shared_ip_count: {shared_ip}
Budget: steps_remaining={steps_left} | tool costs: reverse_image_search=1, analyze_bio=1, check_ip=2
Reply with EXACTLY one token (no punctuation, no explanation):
reverse_image_search (reveal photo_reuse_score)
analyze_bio (reveal bio_template_score)
check_ip (reveal ip_cluster_signal, costs 2 steps)
done (evidence is sufficient, move to flag decision)
"""
DP2_PROMPT = """You must decide whether to FLAG this account as a coordinated fake.
PLATFORM: {platform} | threshold θ*: {threshold:.3f} | fp_penalty: {fp_weight}
ACCOUNT: {account_id} | risk_score: {risk:.3f} | hub_legitimacy: {hub:.2f}
Revealed signals:
photo_reuse_score: {photo}
bio_template_score: {bio}
ip_cluster_signal: {ip}
shared_ip_count: {shared_ip}
Running totals: flagged={n_flagged}/10 | steps_remaining={steps_left}
Guidance:
- Flag ONLY if risk >= threshold and hub_legitimacy < 0.70 (celebrities are hubs).
- High photo_reuse + bio_template or shared_ip_count >= 5 is strong evidence.
- Skip on ambiguous cases when fp_penalty is high.
Reply with EXACTLY one token:
flag (mark this account as a coordinated fake)
skip (leave it alone, move on)
"""
# ---------------------------------------------------------------------------
# Episode log record
# ---------------------------------------------------------------------------
@dataclass
class EpisodeLog:
model: str
platform: str
task: str
seed: int
episode_id: str = ""
threshold: float = 0.0
primary_signal: str = ""
steps_taken: int = 0
inspected: int = 0
tool_calls: Dict[str, int] = field(default_factory=lambda: {
"reverse_image_search": 0, "analyze_bio": 0, "check_ip": 0,
"get_policy": 0, "investigate_network": 0,
})
flagged: int = 0
dp1_calls: int = 0
dp2_calls: int = 0
dp1_invalid: int = 0
dp2_invalid: int = 0
reward: Optional[float] = None
grader_score: Optional[float] = None
final_message: str = ""
wall_seconds: float = 0.0
# ---------------------------------------------------------------------------
# Helpers
# ---------------------------------------------------------------------------
def _find_account(obs, account_id: str):
for a in obs.visible_accounts:
if a.account_id == account_id:
return a
return None
def _render_signal(val) -> str:
if val is None:
return "None"
if isinstance(val, float):
return f"{val:.3f}" if val > 0 else "0.000"
return str(val)
def _parse_dp1(text: str) -> Optional[str]:
t = text.strip().lower().split("\n")[0].strip().strip("'\"`.,")
valid = {"reverse_image_search", "analyze_bio", "check_ip", "done"}
for tok in valid:
if tok in t:
return tok
return None
def _parse_dp2(text: str) -> Optional[str]:
t = text.strip().lower().split("\n")[0].strip().strip("'\"`.,")
if "flag" in t and "unflag" not in t:
return "flag"
if "skip" in t or "no" == t or "keep" in t:
return "skip"
return None
def _seeds_for_platform(seeds: List[int], platform: Optional[str]) -> List[int]:
"""The env assigns platform by seed%2 (even=Instagram, odd=Snapchat).
Offset seeds so the requested platform is actually used."""
if platform is None:
return seeds
p = platform.lower()
if p == "instagram":
return [s if s % 2 == 0 else s + 1 for s in seeds]
if p == "snapchat":
return [s if s % 2 == 1 else s + 1 for s in seeds]
return seeds # unknown platform → env will fall back to its own mapping
def _policy_from_message(msg: str) -> Dict[str, float]:
"""Parse the message returned by GET_POLICY into a dict.
Message format: 'Policy compiled: Platform: X | Threshold: 0.081 | Primary Signal: photo_reuse | FP Penalty: 0.5x | ...'
"""
out = {"threshold": 0.0, "primary_signal": "", "platform": "", "fp_weight": "?"}
try:
body = msg.split("Policy compiled:", 1)[-1]
for part in body.split("|"):
k, _, v = part.strip().partition(":")
k = k.strip().lower()
v = v.strip()
if k == "platform":
out["platform"] = v
elif k == "threshold":
out["threshold"] = float(v)
elif k == "primary signal":
out["primary_signal"] = v
elif k == "fp penalty":
out["fp_weight"] = v
except Exception:
pass
return out
# ---------------------------------------------------------------------------
# Per-account loop (DP1 + DP2)
# ---------------------------------------------------------------------------
def _gather_and_flag(
client: FakeGangEnvClient,
obs,
account_id: str,
policy: Dict,
call_llm: Callable[[str], str],
log: EpisodeLog,
max_dp1_iters: int = 4,
tuples: Optional[List[Dict]] = None,
) -> StepResult:
"""Run DP1 tool-gathering loop then DP2 flag decision for one account.
Returns the latest StepResult (observation reflects any actions taken).
If `tuples` is provided, each LLM decision appends a dict with keys:
prompt, completion, step_reward, decision_type, platform, threshold,
fp_penalty, step_index
"""
last = StepResult(observation=obs, done=False, reward=None, message="")
last.observation = obs
plat = policy.get("platform") or log.platform
fp_w = policy.get("fp_weight", "?")
thr = float(policy.get("threshold", 0.0) or 0.0)
# DP1 loop
for _ in range(max_dp1_iters):
acc = _find_account(last.observation, account_id)
if acc is None:
return last
photo = acc.photo_reuse_score if acc.photo_reuse_score > 0 else None
bio = acc.bio_template_score if acc.bio_template_score > 0 else None
ip_signal = None # ip_cluster_signal is returned only in the step message; treat as unknown until check_ip called
prompt = DP1_PROMPT.format(
platform=policy.get("platform") or log.platform,
primary_signal=policy.get("primary_signal", "?"),
threshold=policy.get("threshold", 0.0),
account_id=account_id,
risk=acc.fake_risk_score,
hub=acc.hub_legitimacy_score,
photo=_render_signal(photo),
bio=_render_signal(bio),
ip=_render_signal(ip_signal),
shared_ip=acc.shared_ip_count,
steps_left=last.observation.steps_remaining,
)
log.dp1_calls += 1
resp = call_llm(prompt)
choice = _parse_dp1(resp)
if choice is None:
log.dp1_invalid += 1
if tuples is not None:
tuples.append({
"prompt": prompt, "completion": resp, "step_reward": 0.0,
"decision_type": "dp1", "platform": plat, "threshold": thr,
"fp_penalty": fp_w, "step_index": log.steps_taken,
})
break
if choice == "done":
if tuples is not None:
tuples.append({
"prompt": prompt, "completion": resp, "step_reward": 0.0,
"decision_type": "dp1", "platform": plat, "threshold": thr,
"fp_penalty": fp_w, "step_index": log.steps_taken,
})
break
atype = {
"reverse_image_search": ActionType.REVERSE_IMAGE_SEARCH,
"analyze_bio": ActionType.ANALYZE_BIO,
"check_ip": ActionType.CHECK_IP,
}[choice]
last = client.step(FakeGangAction(action_type=atype, account_id=account_id))
log.tool_calls[choice] = log.tool_calls.get(choice, 0) + 1
log.steps_taken += 1
if tuples is not None:
tuples.append({
"prompt": prompt, "completion": resp,
"step_reward": float(last.reward) if last.reward is not None else 0.0,
"decision_type": "dp1", "platform": plat, "threshold": thr,
"fp_penalty": fp_w, "step_index": log.steps_taken,
})
if last.done or last.observation.steps_remaining <= 1:
return last
# Stop early if all cheap signals revealed
acc2 = _find_account(last.observation, account_id)
if acc2 and acc2.photo_reuse_score > 0 and acc2.bio_template_score > 0:
break
# DP2
acc = _find_account(last.observation, account_id)
if acc is None:
return last
prompt = DP2_PROMPT.format(
platform=policy.get("platform") or log.platform,
threshold=policy.get("threshold", 0.0),
fp_weight=policy.get("fp_weight", "?"),
account_id=account_id,
risk=acc.fake_risk_score,
hub=acc.hub_legitimacy_score,
photo=_render_signal(acc.photo_reuse_score if acc.photo_reuse_score > 0 else None),
bio=_render_signal(acc.bio_template_score if acc.bio_template_score > 0 else None),
ip=_render_signal(None),
shared_ip=acc.shared_ip_count,
n_flagged=len(last.observation.flagged_ids),
steps_left=last.observation.steps_remaining,
)
log.dp2_calls += 1
resp = call_llm(prompt)
choice = _parse_dp2(resp)
step_reward = 0.0
if choice is None:
log.dp2_invalid += 1
elif choice == "flag":
last = client.step(FakeGangAction(action_type=ActionType.FLAG, account_id=account_id))
log.flagged = len(last.observation.flagged_ids)
step_reward = float(last.reward) if last.reward is not None else 0.0
if tuples is not None:
tuples.append({
"prompt": prompt, "completion": resp, "step_reward": step_reward,
"decision_type": "dp2", "platform": plat, "threshold": thr,
"fp_penalty": fp_w, "step_index": log.steps_taken,
})
return last
# ---------------------------------------------------------------------------
# Single episode
# ---------------------------------------------------------------------------
def _run_episode(
client: FakeGangEnvClient,
model: str,
platform: str,
task: str,
seed: int,
call_llm: Callable[[str], str],
max_accounts_per_episode: int = 15,
collect_tuples: bool = False,
):
"""Run one episode. Returns EpisodeLog by default, or
(EpisodeLog, list[tuple_dict]) when collect_tuples=True."""
log = EpisodeLog(model=model, platform=platform, task=task, seed=seed)
tuples: List[Dict] = [] if collect_tuples else None # type: ignore[assignment]
t0 = time.time()
res = client.reset(task=task, seed=seed)
obs = res.observation
# Deterministic step 0: GET_POLICY
res = client.step(FakeGangAction(action_type=ActionType.GET_POLICY))
log.tool_calls["get_policy"] += 1
policy = _policy_from_message(res.message or res.observation.message)
log.threshold = float(policy.get("threshold", 0.0) or 0.0)
log.primary_signal = str(policy.get("primary_signal", "") or "")
obs = res.observation
investigated: set[str] = set()
processed: set[str] = set()
while not res.done and obs.steps_remaining > 1 and len(processed) < max_accounts_per_episode:
# Candidate pool: suspects first, then visible not-yet-processed, ranked by risk.
candidates = [a for a in obs.visible_accounts if a.account_id not in processed]
suspects = set(obs.suspect_ids)
candidates.sort(
key=lambda a: (a.account_id in suspects, a.fake_risk_score),
reverse=True,
)
if not candidates:
break
acc = candidates[0]
aid = acc.account_id
# Ensure inspected
if aid not in obs.inspected_ids:
res = client.step(FakeGangAction(action_type=ActionType.INSPECT, account_id=aid))
log.steps_taken += 1
log.inspected += 1
obs = res.observation
if res.done or obs.steps_remaining <= 1:
break
# Expand network once for very risky hubs
acc_now = _find_account(obs, aid)
if (
acc_now
and acc_now.fake_risk_score >= 0.80
and aid not in investigated
and obs.steps_remaining >= 5
):
res = client.step(FakeGangAction(action_type=ActionType.INVESTIGATE_NETWORK, account_id=aid))
log.tool_calls["investigate_network"] += 1
log.steps_taken += 2
investigated.add(aid)
obs = res.observation
if res.done or obs.steps_remaining <= 1:
break
# DP1 + DP2 per account
res = _gather_and_flag(client, obs, aid, policy, call_llm, log, tuples=tuples)
obs = res.observation
processed.add(aid)
if res.done:
break
# Final submit (if not already done)
if not res.done:
res = client.step(FakeGangAction(action_type=ActionType.SUBMIT))
log.flagged = len(res.observation.flagged_ids)
log.reward = res.reward
log.final_message = (res.message or "")[:400]
log.episode_id = getattr(res.observation, "episode_id", "") or f"{task}_{seed:03d}_{platform}"
log.wall_seconds = round(time.time() - t0, 2)
# Grader endpoint (optional)
try:
import requests
g = requests.get(f"{client.base_url}/grader", timeout=30).json()
log.grader_score = g.get("score")
except Exception:
pass
if collect_tuples:
# Attach episode-level identifiers to each decision tuple.
for t in tuples:
t["episode_id"] = log.episode_id
t["grader_score"] = log.grader_score
return log, tuples
return log
# ---------------------------------------------------------------------------
# Public entry point
# ---------------------------------------------------------------------------
def run_evaluation(
model_name: str,
call_llm: Callable[[str], str],
platform: str,
base_url: str = "http://localhost:8000",
tasks: Optional[List[str]] = None,
seeds: Optional[List[int]] = None,
out_path: Optional[str] = None,
) -> List[EpisodeLog]:
tasks = tasks or ["easy", "medium", "hard"]
seeds = seeds or [0, 1, 2]
seeds = _seeds_for_platform(seeds, platform)
out_path = out_path or str(
_PARENT / "eval-models" / "results" /
f"{model_name.replace('/', '_')}_{platform.lower()}_results.jsonl"
)
Path(out_path).parent.mkdir(parents=True, exist_ok=True)
print(f"\n{'='*70}")
print(f"Round 2 evaluation | model={model_name} | platform={platform}")
print(f"Target: {base_url} | tasks={tasks} | seeds={seeds}")
print(f"Log: {out_path}")
print(f"{'='*70}")
logs: List[EpisodeLog] = []
client = FakeGangEnvClient(base_url=base_url)
with open(out_path, "w") as f:
for task in tasks:
for seed in seeds:
print(f"\n--- episode: task={task} seed={seed} ---")
try:
log = _run_episode(client, model_name, platform, task, seed, call_llm)
except Exception as e:
print(f" ✗ episode failed: {e}")
log = EpisodeLog(
model=model_name, platform=platform, task=task, seed=seed,
final_message=f"EXCEPTION: {e}",
)
logs.append(log)
f.write(json.dumps(asdict(log)) + "\n")
f.flush()
print(
f" → steps={log.steps_taken} inspected={log.inspected} "
f"flagged={log.flagged} dp1={log.dp1_calls}(bad={log.dp1_invalid}) "
f"dp2={log.dp2_calls}(bad={log.dp2_invalid}) "
f"reward={log.reward} grader={log.grader_score} ({log.wall_seconds}s)"
)
# Summary
print(f"\n{'='*70}")
print("SUMMARY")
print(f"{'='*70}")
rewards = [l.reward for l in logs if l.reward is not None]
graders = [l.grader_score for l in logs if l.grader_score is not None]
if rewards:
print(f" mean reward: {sum(rewards)/len(rewards):.4f} (n={len(rewards)})")
if graders:
print(f" mean grader: {sum(graders)/len(graders):.4f} (n={len(graders)})")
total_dp1 = sum(l.dp1_calls for l in logs)
total_dp2 = sum(l.dp2_calls for l in logs)
bad_dp1 = sum(l.dp1_invalid for l in logs)
bad_dp2 = sum(l.dp2_invalid for l in logs)
print(f" DP1 calls: {total_dp1} (invalid: {bad_dp1})")
print(f" DP2 calls: {total_dp2} (invalid: {bad_dp2})")
print(f" Logged to: {out_path}")
return logs
# ---------------------------------------------------------------------------
# Standard CLI argument parser (reused by every model shim)
# ---------------------------------------------------------------------------
def build_cli():
import argparse
parser = argparse.ArgumentParser(description="Round 2 eval runner")
parser.add_argument("--url", default=os.getenv("API_BASE_URL_ENV", "http://localhost:8000"),
help="Environment server URL")
parser.add_argument("--platform", default="Instagram", help="Platform name (Instagram/Snapchat/...)")
parser.add_argument("--tasks", nargs="+", default=["easy", "medium", "hard"])
parser.add_argument("--seeds", nargs="+", type=int, default=[0, 1, 2])
parser.add_argument("--out", default=None, help="Output JSONL path")
return parser
|