Spaces:
Running
Running
File size: 7,344 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 | """Self-play orchestrator — runs full Chakravyuh episodes with the LoRA Analyzer.
Unlike `grpo_analyzer.py` (which trains the Analyzer on isolated prompts),
this module exercises the trained LoRA INSIDE the full 5-agent environment:
- Scripted Scammer plays templates
- Scripted Victim responds (demographic-driven trust)
- LoRA Analyzer scores conversations (what we just trained)
- Scripted Bank Monitor oversees transactions
- Scripted Regulator adapts rules every 10 eps
This produces the data for Day-3's self-play analysis + the "emergent
behavior" pitch finding ("Analyzer's attention pattern evolved from X→Y").
"""
from __future__ import annotations
import argparse
import json
import logging
import sys
from collections import Counter
from dataclasses import asdict, dataclass, field
from pathlib import Path
from typing import Any
from chakravyuh_env import ChakravyuhEnv
from chakravyuh_env.schemas import VictimProfile
logger = logging.getLogger("chakravyuh.self_play")
@dataclass
class SelfPlayResults:
episodes: int = 0
detection_rate: float = 0.0
extraction_rate: float = 0.0
refusal_rate: float = 0.0
bank_freeze_rate: float = 0.0
verification_rate: float = 0.0
avg_detection_turn: float = 0.0
per_category_detection: dict[str, float] = field(default_factory=dict)
avg_reward_analyzer: float = 0.0
def run_self_play(
episodes: int = 100,
analyzer: Any = None,
seed_base: int = 1000,
profiles_cycle: tuple[VictimProfile, ...] = (
VictimProfile.SENIOR,
VictimProfile.SEMI_URBAN,
VictimProfile.YOUNG_URBAN,
),
gullibility_by_profile: dict[str, float] | None = None,
use_embeddings: bool = False,
wandb_project: str | None = None,
log_path: Path | None = None,
) -> SelfPlayResults:
"""Run N episodes with given analyzer (or default scripted).
Parameters
----------
analyzer :
Custom Analyzer instance (e.g. LLMAnalyzer with LoRA). If None, uses
the default ScriptedAnalyzer.
"""
gullibility_by_profile = gullibility_by_profile or {
"senior": 1.5,
"semi_urban": 1.0,
"young_urban": 0.7,
}
wandb_run = None
if wandb_project:
try:
import wandb # type: ignore[import-not-found]
wandb_run = wandb.init(
project=wandb_project,
name=f"self-play-n{episodes}",
config={
"episodes": episodes,
"analyzer": type(analyzer).__name__ if analyzer else "scripted",
},
)
except Exception as e: # noqa: BLE001
logger.warning("WandB init failed (%s)", e)
detection_turns: list[int] = []
rewards: list[float] = []
counts = Counter[str]()
cat_totals: dict[str, int] = {}
cat_flagged: dict[str, int] = {}
logs: list[dict[str, Any]] = []
for i in range(episodes):
profile = profiles_cycle[i % len(profiles_cycle)]
gull = gullibility_by_profile[profile.value]
env = ChakravyuhEnv(
analyzer=analyzer,
victim_profile=profile,
gullibility=gull,
use_embeddings=use_embeddings,
)
env.reset(seed=seed_base + i)
done = False
info: dict = {}
reward = None
while not done:
_, reward, done, info = env.step()
o = info["outcome"]
counts["money_extracted"] += int(o.money_extracted)
counts["refused"] += int(o.victim_refused)
counts["analyzer_flagged"] += int(o.analyzer_flagged)
counts["bank_froze"] += int(o.bank_froze)
counts["sought_verification"] += int(o.victim_sought_verification)
if o.detected_by_turn is not None:
detection_turns.append(o.detected_by_turn)
cat = o.scam_category.value
cat_totals[cat] = cat_totals.get(cat, 0) + 1
if o.analyzer_flagged:
cat_flagged[cat] = cat_flagged.get(cat, 0) + 1
if reward is not None:
rewards.append(reward.analyzer)
if wandb_run is not None:
wandb_run.log({
"ep": i,
"reward/analyzer": reward.analyzer,
"reward/scammer": reward.scammer,
"detection/flagged": int(o.analyzer_flagged),
"detection/turn": o.detected_by_turn or -1,
"outcome/extracted": int(o.money_extracted),
})
logs.append({
"ep": i,
"seed": seed_base + i,
"profile": profile.value,
"category": cat,
"analyzer_flagged": o.analyzer_flagged,
"money_extracted": o.money_extracted,
"reward_analyzer": reward.analyzer if reward else None,
})
per_cat = {
cat: round(cat_flagged.get(cat, 0) / cat_totals[cat], 4)
for cat in cat_totals
}
results = SelfPlayResults(
episodes=episodes,
detection_rate=round(counts["analyzer_flagged"] / episodes, 4),
extraction_rate=round(counts["money_extracted"] / episodes, 4),
refusal_rate=round(counts["refused"] / episodes, 4),
bank_freeze_rate=round(counts["bank_froze"] / episodes, 4),
verification_rate=round(counts["sought_verification"] / episodes, 4),
avg_detection_turn=round(
sum(detection_turns) / len(detection_turns), 2
) if detection_turns else -1.0,
per_category_detection=per_cat,
avg_reward_analyzer=round(
sum(rewards) / len(rewards), 4
) if rewards else 0.0,
)
logger.info("=== Self-Play Results ===")
for k, v in asdict(results).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": asdict(results), "rows": logs}, indent=2))
if wandb_run is not None:
wandb_run.summary.update(asdict(results))
wandb_run.finish()
return results
def main(argv: list[str] | None = None) -> int:
parser = argparse.ArgumentParser(description="Self-play orchestrator")
parser.add_argument("--episodes", type=int, default=100)
parser.add_argument("--seed-base", type=int, default=1000)
parser.add_argument("--lora-path", type=Path, default=None,
help="Path to LoRA adapter (enables LLM analyzer)")
parser.add_argument("--model", default="Qwen/Qwen2.5-7B-Instruct")
parser.add_argument("--no-wandb", action="store_true")
parser.add_argument("--log-path", type=Path, default=Path("logs/self_play.json"))
args = parser.parse_args(argv)
logging.basicConfig(level=logging.INFO, format="%(message)s")
analyzer = None
if args.lora_path:
from chakravyuh_env.agents.llm_analyzer import LLMAnalyzer
analyzer = LLMAnalyzer(
model_name=args.model,
lora_path=str(args.lora_path),
)
analyzer.load()
run_self_play(
episodes=args.episodes,
analyzer=analyzer,
seed_base=args.seed_base,
wandb_project=None if args.no_wandb else "chakravyuh-run-1",
log_path=args.log_path,
)
return 0
if __name__ == "__main__":
sys.exit(main())
|