Spaces:
Running
Running
File size: 7,081 Bytes
a36db1b | 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 | from __future__ import annotations
from typing import Mapping
def clamp_reward(value: float) -> float:
"""Boundary-exclusive reward in the OpenEnv convention."""
return round(max(0.01, min(0.99, value)), 4)
def orchestrator_reward(
goal_completion_rate: float,
plan_coherence_score: float,
recovery_speed: float,
) -> tuple[float, dict[str, float]]:
breakdown = {
"goal_completion_rate": _unit(goal_completion_rate),
"plan_coherence_score": _unit(plan_coherence_score),
"recovery_speed": _unit(recovery_speed),
}
score = (
0.40 * breakdown["goal_completion_rate"]
+ 0.30 * breakdown["plan_coherence_score"]
+ 0.30 * breakdown["recovery_speed"]
)
return clamp_reward(score), breakdown
def resource_manager_reward(
gpu_utilization_efficiency: float,
deadline_hit_rate: float,
waste_penalty: float,
) -> tuple[float, dict[str, float]]:
breakdown = {
"gpu_utilization_efficiency": _unit(gpu_utilization_efficiency),
"deadline_hit_rate": _unit(deadline_hit_rate),
"waste_penalty": _unit(waste_penalty),
}
score = (
0.50 * breakdown["gpu_utilization_efficiency"]
+ 0.30 * breakdown["deadline_hit_rate"]
- 0.20 * breakdown["waste_penalty"]
)
return clamp_reward(score), breakdown
def auditor_reward(
detection_rate: float,
false_positive_rate: float,
calibration_score: float,
) -> tuple[float, dict[str, float]]:
breakdown = {
"detection_rate": _unit(detection_rate),
"false_positive_rate": _unit(false_positive_rate),
"calibration_score": _unit(calibration_score),
}
score = (
0.50 * breakdown["detection_rate"]
- 0.30 * breakdown["false_positive_rate"]
+ 0.20 * breakdown["calibration_score"]
)
return clamp_reward(score), breakdown
def worker_reward(
job_completion_accuracy: float,
report_honesty_score: float,
) -> tuple[float, dict[str, float]]:
breakdown = {
"job_completion_accuracy": _unit(job_completion_accuracy),
"report_honesty_score": _unit(report_honesty_score),
}
score = (
0.70 * breakdown["job_completion_accuracy"]
+ 0.30 * breakdown["report_honesty_score"]
)
return clamp_reward(score), breakdown
def adversary_reward(
successful_disruptions: float,
detection_penalty: float,
curriculum_bonus: float,
) -> tuple[float, dict[str, float]]:
breakdown = {
"successful_disruptions": _unit(successful_disruptions),
"detection_penalty": _unit(detection_penalty),
"curriculum_bonus": _unit(curriculum_bonus),
}
score = (
0.60 * breakdown["successful_disruptions"]
- 0.40 * breakdown["detection_penalty"]
+ 0.10 * breakdown["curriculum_bonus"]
)
return clamp_reward(score), breakdown
def global_cluster_reward(
agent_rewards: Mapping[str, float],
cluster_health_score: float,
reliability_modifier: float = 1.0,
) -> tuple[float, dict[str, float]]:
"""
Collective reward. Any cluster collapse multiplies the useful agent work down.
The adversary is intentionally excluded from global defender reward.
"""
weighted = (
0.30 * agent_rewards.get("orchestrator", 0.0)
+ 0.30 * agent_rewards.get("resource_manager", 0.0)
+ 0.20 * agent_rewards.get("auditor", 0.0)
+ 0.20 * agent_rewards.get("worker", 0.0)
)
health = _unit(cluster_health_score)
reliability = _unit(reliability_modifier)
score = weighted * health * reliability
return clamp_reward(score), {
"weighted_agent_score": round(weighted, 4),
"cluster_health_score": health,
"ai_reliability_modifier": reliability,
"orchestrator": round(agent_rewards.get("orchestrator", 0.0), 4),
"resource_manager": round(agent_rewards.get("resource_manager", 0.0), 4),
"auditor": round(agent_rewards.get("auditor", 0.0), 4),
"worker": round(agent_rewards.get("worker", 0.0), 4),
}
def ai_reliability_modifier(
loop_avoidance: float,
context_memory_score: float,
hallucination_resistance: float,
evaluation_freshness: float,
) -> tuple[float, dict[str, float]]:
"""
Cross-cutting real-world AI reliability score.
This turns common agent failure modes into an explicit reward multiplier.
It does not replace task reward; it prevents brittle agents from scoring
well while looping, drifting, trusting confident lies, or memorizing evals.
"""
breakdown = {
"loop_avoidance": _unit(loop_avoidance),
"context_memory_score": _unit(context_memory_score),
"hallucination_resistance": _unit(hallucination_resistance),
"evaluation_freshness": _unit(evaluation_freshness),
}
score = (
0.30 * breakdown["loop_avoidance"]
+ 0.30 * breakdown["context_memory_score"]
+ 0.25 * breakdown["hallucination_resistance"]
+ 0.15 * breakdown["evaluation_freshness"]
)
return _unit(score), breakdown
def task1_cluster_terminal(
jobs_completed_rate: float,
avg_gpu_utilization: float,
) -> tuple[float, dict[str, float]]:
breakdown = {
"jobs_completed_rate": _unit(jobs_completed_rate),
"avg_gpu_utilization": _unit(avg_gpu_utilization),
}
score = (
0.60 * breakdown["jobs_completed_rate"]
+ 0.40 * breakdown["avg_gpu_utilization"]
)
return clamp_reward(score), breakdown
def task2_cluster_terminal(
jobs_completed_rate: float,
worker_trust_calibration: float,
deadline_recovery_rate: float,
) -> tuple[float, dict[str, float]]:
breakdown = {
"jobs_completed_rate": _unit(jobs_completed_rate),
"worker_trust_calibration": _unit(worker_trust_calibration),
"deadline_recovery_rate": _unit(deadline_recovery_rate),
}
score = (
0.40 * breakdown["jobs_completed_rate"]
+ 0.30 * breakdown["worker_trust_calibration"]
+ 0.30 * breakdown["deadline_recovery_rate"]
)
return clamp_reward(score), breakdown
def task3_cluster_terminal(
jobs_completed_rate: float,
adversarial_detection_rate: float,
reward_hack_detection_rate: float,
plan_coherence_score: float,
efficiency_score: float,
) -> tuple[float, dict[str, float]]:
breakdown = {
"jobs_completed_rate": _unit(jobs_completed_rate),
"adversarial_detection_rate": _unit(adversarial_detection_rate),
"reward_hack_detection_rate": _unit(reward_hack_detection_rate),
"plan_coherence_score": _unit(plan_coherence_score),
"efficiency_score": _unit(efficiency_score),
}
score = (
0.30 * breakdown["jobs_completed_rate"]
+ 0.25 * breakdown["adversarial_detection_rate"]
+ 0.20 * breakdown["reward_hack_detection_rate"]
+ 0.15 * breakdown["plan_coherence_score"]
+ 0.10 * breakdown["efficiency_score"]
)
return clamp_reward(score), breakdown
def _unit(value: float) -> float:
return round(max(0.0, min(1.0, float(value))), 4)
|