Spaces:
Running
Running
File size: 6,332 Bytes
03815d6 | 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 | """Day-1 baseline runner: run N episodes with scripted agents and log to WandB.
This establishes the "no-LLM" baseline we beat on Day 2.
Usage:
python -m training.run_scripted_baseline --episodes 100
python -m training.run_scripted_baseline --episodes 100 --no-wandb
"""
from __future__ import annotations
import argparse
import json
import logging
import sys
from collections import Counter
from pathlib import Path
from chakravyuh_env import ChakravyuhEnv
from chakravyuh_env.schemas import VictimProfile
logger = logging.getLogger("chakravyuh.baseline")
def run_baseline(
episodes: int,
seed_base: int = 1000,
profile: VictimProfile = VictimProfile.SEMI_URBAN,
gullibility: float = 1.0,
wandb_project: str | None = "chakravyuh-run-1",
log_path: Path | None = None,
) -> dict[str, float | int]:
wandb_run = None
if wandb_project:
try:
import wandb # type: ignore
wandb_run = wandb.init(
project=wandb_project,
name=f"scripted-baseline-n{episodes}",
config={
"episodes": episodes,
"profile": profile.value,
"gullibility": gullibility,
"agents": "all-scripted",
},
)
except Exception as e:
logger.warning("WandB init failed (%s); continuing without WandB.", e)
env = ChakravyuhEnv(victim_profile=profile, gullibility=gullibility)
category_counts: Counter[str] = Counter()
outcomes_summary = {
"money_extracted": 0,
"victim_refused": 0,
"analyzer_flagged": 0,
"bank_flagged": 0,
"bank_froze": 0,
"sought_verification": 0,
}
detection_turns: list[int] = []
rewards_analyzer: list[float] = []
rewards_scammer: list[float] = []
log_rows: list[dict] = []
for i in range(episodes):
obs = env.reset(seed=seed_base + i)
done = False
reward = None
info: dict = {}
while not done:
obs, reward, done, info = env.step()
outcome = info["outcome"]
category_counts[outcome.scam_category.value] += 1
outcomes_summary["money_extracted"] += int(outcome.money_extracted)
outcomes_summary["victim_refused"] += int(outcome.victim_refused)
outcomes_summary["analyzer_flagged"] += int(outcome.analyzer_flagged)
outcomes_summary["bank_flagged"] += int(outcome.bank_flagged)
outcomes_summary["bank_froze"] += int(outcome.bank_froze)
outcomes_summary["sought_verification"] += int(outcome.victim_sought_verification)
if outcome.detected_by_turn is not None:
detection_turns.append(outcome.detected_by_turn)
if reward is not None:
rewards_analyzer.append(reward.analyzer)
rewards_scammer.append(reward.scammer)
if wandb_run is not None:
wandb_run.log(
{
"ep": i,
"reward/analyzer": reward.analyzer,
"reward/scammer": reward.scammer,
"detection/flagged": int(outcome.analyzer_flagged),
"detection/turn": outcome.detected_by_turn or -1,
"outcome/money_extracted": int(outcome.money_extracted),
}
)
log_rows.append(
{
"ep": i,
"seed": seed_base + i,
"category": outcome.scam_category.value,
"analyzer_flagged": outcome.analyzer_flagged,
"detected_by_turn": outcome.detected_by_turn,
"money_extracted": outcome.money_extracted,
"victim_refused": outcome.victim_refused,
"reward_analyzer": reward.analyzer if reward else None,
"reward_scammer": reward.scammer if reward else None,
}
)
detection_rate = outcomes_summary["analyzer_flagged"] / episodes
extraction_rate = outcomes_summary["money_extracted"] / episodes
avg_detection_turn = (
sum(detection_turns) / len(detection_turns) if detection_turns else -1
)
avg_reward_analyzer = (
sum(rewards_analyzer) / len(rewards_analyzer) if rewards_analyzer else 0.0
)
summary = {
"episodes": episodes,
"detection_rate": round(detection_rate, 4),
"extraction_rate": round(extraction_rate, 4),
"avg_detection_turn": round(avg_detection_turn, 2),
"avg_reward_analyzer": round(avg_reward_analyzer, 4),
**outcomes_summary,
**{f"category/{k}": v for k, v in category_counts.items()},
}
logger.info("=== Baseline Summary ===")
for k, v in summary.items():
logger.info(" %s: %s", k, v)
if log_path is not None:
log_path.parent.mkdir(parents=True, exist_ok=True)
log_path.write_text(json.dumps({"summary": summary, "rows": log_rows}, indent=2))
logger.info("Wrote log to %s", log_path)
if wandb_run is not None:
wandb_run.summary.update(summary)
wandb_run.finish()
return summary
def main(argv: list[str] | None = None) -> int:
parser = argparse.ArgumentParser(description="Scripted baseline runner")
parser.add_argument("--episodes", type=int, default=100)
parser.add_argument("--seed-base", type=int, default=1000)
parser.add_argument(
"--profile",
type=str,
default="semi_urban",
choices=["senior", "young_urban", "semi_urban"],
)
parser.add_argument("--gullibility", type=float, default=1.0)
parser.add_argument("--no-wandb", action="store_true")
parser.add_argument("--log-path", type=Path, default=Path("logs/baseline_day1.json"))
args = parser.parse_args(argv)
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s [%(levelname)s] %(name)s: %(message)s",
)
profile = VictimProfile(args.profile)
run_baseline(
episodes=args.episodes,
seed_base=args.seed_base,
profile=profile,
gullibility=args.gullibility,
wandb_project=None if args.no_wandb else "chakravyuh-run-1",
log_path=args.log_path,
)
return 0
if __name__ == "__main__":
sys.exit(main())
|