Spaces:
Sleeping
Sleeping
File size: 9,845 Bytes
15503f9 599c9bd 15503f9 599c9bd 15503f9 | 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 | """
Base reward infrastructure β data classes, calculators, and transforms.
Merged from the shared repo-level modules into a self-contained file:
- Episode-level: RewardCalculator (custom mode)
- Per-step: StepRewardTransform + OpenEnvRewardCalculator (openenv mode)
Scoring formula (both modes):
total = 0.25 * quality/structural + 0.15 * efficiency + 0.60 * ground_truth + penalty
Usage:
from rewards.base import RewardCalculator, Scenario, EpisodeLog
calculator = RewardCalculator()
breakdown = calculator.calculate(episode, scenario, outcome_results)
"""
from dataclasses import dataclass, field
from typing import Any, Dict, List, Optional, Set
from openenv.core.env_server.interfaces import Transform
from openenv.core.env_server.mcp_types import CallToolObservation
from openenv.core.env_server.types import Observation
# ββ Data Classes ββ
@dataclass
class StepLog:
"""Record of a single tool call made by the agent."""
tool_name: str
arguments: Dict[str, Any]
success: bool
result: Any = None
error: Optional[str] = None
timestamp: Optional[str] = None
elapsed: float = 0.0
@dataclass
class EpisodeLog:
"""Record of all tool calls in one episode."""
steps: List[StepLog] = field(default_factory=list)
def add_step(
self,
tool_name: str,
arguments: Dict[str, Any],
success: bool,
result: Any = None,
error: Optional[str] = None,
timestamp: Optional[str] = None,
elapsed: float = 0.0,
) -> None:
self.steps.append(
StepLog(
tool_name=tool_name,
arguments=arguments,
success=success,
result=result,
error=error,
timestamp=timestamp,
elapsed=elapsed,
)
)
@property
def tools_used(self) -> List[str]:
return [s.tool_name for s in self.steps]
@property
def tools_used_set(self) -> Set[str]:
return set(self.tools_used)
@dataclass
class Scenario:
"""Definition of a task for the agent."""
id: str
prompt: str
expected_tools: List[str]
max_steps: int
outcome_checks: List[Dict[str, Any]]
@dataclass
class RewardBreakdown:
"""Detailed reward breakdown β useful for debugging and logging."""
structural: float = 0.0
ground_truth: float = 0.0
efficiency: float = 0.0
penalty: float = 0.0
total: float = 0.0
details: Dict[str, Any] = field(default_factory=dict)
def summary(self) -> str:
mode = self.details.get("reward_mode", "custom")
qual_label = "Quality" if mode == "openenv" else "Structural"
lines = [
f" {qual_label + ':':14s}{self.structural:.2f} (weight 0.25)",
f" Efficiency: {self.efficiency:.2f} (weight 0.15)",
f" Ground Truth: {self.ground_truth:.2f} (weight 0.60)",
]
if self.penalty < 0:
lines.append(f" Penalty: {self.penalty:.2f} (hallucination)")
lines.append(f" ββββββββββββββββββββββββ")
lines.append(f" TOTAL: {self.total:.2f}")
return "\n".join(lines)
# ββ Episode-Level Reward Calculator (custom mode) ββ
class RewardCalculator:
"""
Computes episode-level reward from logs + scenario + verification results.
Weights: structural (0.25), ground_truth (0.60), efficiency (0.15).
"""
def __init__(
self,
w_structural: float = 0.25,
w_ground_truth: float = 0.60,
w_efficiency: float = 0.15,
):
self.w_structural = w_structural
self.w_ground_truth = w_ground_truth
self.w_efficiency = w_efficiency
def calculate(
self,
episode: EpisodeLog,
scenario: Scenario,
outcome_results: List[float],
) -> RewardBreakdown:
breakdown = RewardBreakdown()
breakdown.structural = self._structural_score(episode, scenario)
breakdown.ground_truth = self._ground_truth_score(outcome_results)
breakdown.efficiency = self._efficiency_score(episode, scenario)
breakdown.penalty = self._hallucination_penalty(episode, outcome_results)
breakdown.total = (
self.w_structural * breakdown.structural
+ self.w_ground_truth * breakdown.ground_truth
+ self.w_efficiency * breakdown.efficiency
+ breakdown.penalty
)
breakdown.total = max(-1.0, min(1.0, breakdown.total))
breakdown.details = {
"tools_expected": scenario.expected_tools,
"tools_used": episode.tools_used,
"outcome_checks_score_sum": sum(outcome_results),
"outcome_checks_total": len(outcome_results),
"outcome_checks_avg": sum(outcome_results) / len(outcome_results) if outcome_results else 0.0,
"steps_taken": len(episode.steps),
"max_steps": scenario.max_steps,
}
return breakdown
def _structural_score(self, episode: EpisodeLog, scenario: Scenario) -> float:
if not episode.steps:
return 0.0
expected = set(scenario.expected_tools)
used = episode.tools_used_set
intersection = expected & used
precision = len(intersection) / len(used) if used else 0.0
recall = len(intersection) / len(expected) if expected else 0.0
f1 = (
2 * precision * recall / (precision + recall)
if (precision + recall) > 0
else 0.0
)
success_rate = sum(1 for s in episode.steps if s.success) / len(episode.steps)
unexpected_calls = sum(
1 for s in episode.steps if s.tool_name not in expected
)
unexpected_ratio = unexpected_calls / len(episode.steps)
return max(0.0, 0.6 * f1 + 0.4 * success_rate - unexpected_ratio * 0.3)
def _ground_truth_score(self, outcome_results: List[float]) -> float:
if not outcome_results:
return 0.0
return sum(outcome_results) / len(outcome_results)
def _efficiency_score(self, episode: EpisodeLog, scenario: Scenario) -> float:
if not episode.steps:
return 0.0
return max(0.0, 1.0 - len(episode.steps) / scenario.max_steps)
def _hallucination_penalty(
self, episode: EpisodeLog, outcome_results: List[float]
) -> float:
if not episode.steps or not outcome_results:
return 0.0
all_calls_succeeded = all(s.success for s in episode.steps)
pass_rate = sum(outcome_results) / len(outcome_results)
if all_calls_succeeded and pass_rate == 0.0:
return -0.5
if all_calls_succeeded and pass_rate < 0.3:
return -0.2
return 0.0
# ββ Per-Step Reward Transform (openenv mode) ββ
class StepRewardTransform(Transform):
"""
Gym-agnostic per-step reward transform.
Sets observation.reward based on tool call success/failure.
Subclass for gym-specific logic (see transforms.py).
"""
def __call__(self, observation: Observation) -> Observation:
reward = self._compute_reward(observation)
observation.reward = reward
return observation
def _compute_reward(self, observation: Observation) -> float:
if isinstance(observation, CallToolObservation):
if observation.error is not None:
return -0.5
return 1.0
return 0.0
class OpenEnvRewardCalculator:
"""
Combines per-step transform rewards with ground truth verification.
Used as the alternative to RewardCalculator when --reward-mode openenv.
Quality is sign-based: only the sign of per-step rewards matters
(positive = productive, negative = harmful, zero = neutral).
"""
def __init__(
self,
w_quality: float = 0.25,
w_efficiency: float = 0.15,
w_ground_truth: float = 0.60,
):
self.w_quality = w_quality
self.w_efficiency = w_efficiency
self.w_ground_truth = w_ground_truth
def calculate(
self,
step_rewards: List[float],
outcome_results: List[bool],
max_steps: int = 0,
actual_steps: int = 0,
) -> RewardBreakdown:
productive = sum(1 for r in step_rewards if r > 0)
harmful = sum(1 for r in step_rewards if r < 0)
active = productive + harmful
quality = productive / active if active > 0 else 0.0
if max_steps > 0 and actual_steps > 0:
efficiency = max(0.0, 1.0 - actual_steps / max_steps)
else:
efficiency = 0.0
gt_score = sum(outcome_results) / len(outcome_results) if outcome_results else 0.0
penalty = 0.0
if step_rewards and outcome_results:
no_harmful = all(r >= 0 for r in step_rewards)
if no_harmful and gt_score == 0.0:
penalty = -0.5
elif no_harmful and gt_score < 0.3:
penalty = -0.2
total = (
self.w_quality * quality
+ self.w_efficiency * efficiency
+ self.w_ground_truth * gt_score
+ penalty
)
total = max(-1.0, min(1.0, total))
return RewardBreakdown(
structural=quality,
ground_truth=gt_score,
efficiency=efficiency,
penalty=penalty,
total=total,
details={
"reward_mode": "openenv",
"productive_steps": productive,
"harmful_steps": harmful,
"neutral_steps": len(step_rewards) - active,
"actual_steps": actual_steps,
"max_steps": max_steps,
},
)
|