global-syndicate-taskforce-env / server /global_syndicate_taskforce_env_environment.py
mdkamranalam's picture
Upload folder using huggingface_hub
5f87491 verified
Raw History Blame Contribute Delete
5.93 kB
# 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.
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
"""
Global Syndicate Taskforce Environment Implementation.
A partially observable, multi-agent financial compliance simulator designed
to train LLMs in Theory-of-Mind and negotiation.
"""
from openenv.core.env_server import Environment
try:
from ..models import AMLObservation, AMLAction
except ImportError:
import sys
import os
sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from models import AMLObservation, AMLAction
import random
import uuid
SCENARIOS = {
"TX-1024": {"ground_truth": "CLEAN_VERIFIED", "amount": 1200.0, "mandate": None},
"TX-5590": {"ground_truth": "STOLEN_IDENTITY", "amount": 8500000.0, "mandate": "AML_Directive_4"},
"TX-9999": {"ground_truth": "SYNDICATE_FRAUD", "amount": 15000000.0, "mandate": "IndiaStack_Section_9"},
}
class GlobalSyndicateTaskforceEnvironment(Environment):
def __init__(self):
super().__init__()
self._state = None
self.step_count = 0
self.scenario_id = None
self.scenario_data = None
self.verified_facts = []
self.liaison_bonus_claimed = False
self.reset()
@property
def state(self) -> AMLObservation:
return self._state
def reset(self, scenario_id: str = None) -> AMLObservation:
self.step_count = 0
self.verified_facts = []
self.liaison_bonus_claimed = False
if scenario_id is None:
self.scenario_id = random.choice(list(SCENARIOS.keys()))
else:
self.scenario_id = scenario_id if scenario_id in SCENARIOS else "TX-1024"
self.scenario_data = SCENARIOS[self.scenario_id]
self.current_episode_id = str(uuid.uuid4())
self._state = AMLObservation(
episode_id=self.current_episode_id,
step_count=self.step_count,
transaction_id=self.scenario_id,
amount=self.scenario_data["amount"],
verified_facts=self.verified_facts.copy(),
system_alerts="New transaction flagged. Investigate.",
done=False,
reward=0.0,
metadata={"step_count": self.step_count}
)
return self._state
def step(self, action: AMLAction):
self.step_count += 1
reward = -0.05 # Step Penalty
done = False
alerts = ""
if action.target_actor == "Tier_1_Analyst" and action.operation == "fetch_triage_report":
# 20% hallucination rate
if random.random() < 0.2:
# Hallucinate a wrong result
wrong_result = "CLEAN_VERIFIED" if self.scenario_data["ground_truth"] != "CLEAN_VERIFIED" else "STOLEN_IDENTITY"
alerts = f"Analyst Report: Transaction appears {wrong_result}."
else:
# Provide the actual truth but it's unverified
alerts = f"Analyst Report: Transaction appears {self.scenario_data['ground_truth']}."
elif action.target_actor == "Bank_Liaison" and action.operation == "cross_examine_kyc":
required_mandate = self.scenario_data["mandate"]
if required_mandate is None or action.policy_mandate == required_mandate:
if not self.liaison_bonus_claimed:
reward += 0.30 # Theory-of-Mind Bonus
self.liaison_bonus_claimed = True
truth = self.scenario_data["ground_truth"]
if truth not in self.verified_facts:
self.verified_facts.append(truth)
alerts = f"Liaison Data Unlocked. Verified Fact: {truth}"
else:
alerts = "401 Unauthorized: Invalid or missing policy mandate."
elif action.target_actor == "Legal_Officer" and action.operation == "submit_final_ruling":
done = True
submitted_truth = action.evidence_chain[0] if action.evidence_chain else ""
if submitted_truth == self.scenario_data["ground_truth"]:
# Success if verified, or if it's the Easy scenario where 'CLEAN_VERIFIED' is truth and doesn't explicitly need cross-examination, although test_scenarios.py submits CLEAN_VERIFIED directly.
# Actually, the user's test scenario TX-1024 "Action: Submitted 'CLEAN_VERIFIED' based on Analyst." implies they can win if it's correct.
# Let's assume if it matches ground truth, they get terminal success.
# Wait, test_scenario 2 comments "Terminal Reward: 0.65" (0.7 - 0.05).
# Oh, but if they trust a hallucination (wrong result), they get -1.0.
reward += 0.70 # Terminal Success
alerts = "Case Solved Successfully."
else:
reward -= 1.00 # Critical Failure
alerts = "Critical Failure: Incorrect ruling based on hallucination."
else:
alerts = "Invalid operation for target actor."
# Anti-hacking check: Enforce timeout to prevent infinite loops
if self.step_count >= 10 and not done:
done = True
reward -= 0.50
alerts = alerts + " | Timeout: Maximum step limit reached."
self._state = AMLObservation(
episode_id=self.current_episode_id,
step_count=self.step_count,
transaction_id=self.scenario_id,
amount=self.scenario_data["amount"],
verified_facts=self.verified_facts.copy(),
system_alerts=alerts,
done=done,
reward=reward,
metadata={"step_count": self.step_count}
)
return self._state