Spaces:
Sleeping
Sleeping
File size: 4,207 Bytes
ff293b1 ee21104 ff293b1 ee21104 ff293b1 ee21104 ff293b1 ee21104 ff293b1 ee21104 ff293b1 ee21104 ff293b1 ee21104 ff293b1 | 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 | """
Public trajectory graders for OpenEnv Phase 2 / HF deep validation.
These are **episode-level** scores (strictly inside (0, 1)), separate from per-step
rewards in `server/reward.py`. The hackathon validator reads `openenv.yaml`
`tasks[].grader` and calls these functions with trajectory dicts.
"""
from __future__ import annotations
import math
from typing import List
STRICT_MIN = 0.01
STRICT_MAX = 0.99
def _bounded(value: float) -> float:
try:
v = round(float(value), 4)
except (TypeError, ValueError):
return 0.5
if not math.isfinite(v):
return 0.5
return min(max(v, STRICT_MIN), STRICT_MAX)
def _as_reward_list(trajectory: dict | None) -> List[float]:
payload = trajectory or {}
if not isinstance(payload, dict):
return []
rewards = payload.get("rewards")
if isinstance(rewards, list) and rewards:
out: List[float] = []
for r in rewards:
try:
rv = float(r)
except (TypeError, ValueError):
continue
if math.isfinite(rv):
out.append(rv)
return out
if "score" in payload:
try:
v = float(payload["score"])
return [v] if math.isfinite(v) else []
except (TypeError, ValueError):
return []
reward = payload.get("reward")
if isinstance(reward, dict) and "total" in reward:
try:
v = float(reward["total"])
return [v] if math.isfinite(v) else []
except (TypeError, ValueError):
return []
if reward is not None:
try:
v = float(reward)
return [v] if math.isfinite(v) else []
except (TypeError, ValueError):
return []
return []
def _profile(reward: float) -> str:
if reward <= 0.05:
return "unsafe_miss"
if reward <= 0.20:
return "bad_call"
if reward < 0.50:
return "weak"
if reward < 0.80:
return "workable"
if reward < 0.95:
return "strong"
return "expert"
def _score_episode(
rewards: List[float],
*,
miss_cost: float,
overcall_cost: float,
stability_gain: float,
expertise_gain: float,
) -> float:
if not rewards:
return _bounded(0.5)
labels = [_profile(r) for r in rewards]
mean_r = sum(rewards) / len(rewards)
n = len(rewards)
miss = labels.count("unsafe_miss")
bad = labels.count("bad_call")
weak = labels.count("weak")
strong = labels.count("strong") + labels.count("expert")
expert = labels.count("expert")
downward = (
min(miss * miss_cost, 0.35)
+ min(bad * overcall_cost, 0.15)
+ min(weak * 0.015, 0.06)
)
upward = 0.0
if strong / n >= 0.80:
upward += stability_gain
if expert / n >= 0.60:
upward += expertise_gain
return _bounded(mean_r - downward + upward)
def phase2_core_grader(trajectory: dict | None = None) -> float:
"""Easy tier — dense default inbox (scenarios/phase2_core.json)."""
return _score_episode(
_as_reward_list(trajectory),
miss_cost=0.12,
overcall_cost=0.03,
stability_gain=0.05,
expertise_gain=0.01,
)
def monday_morning_grader(trajectory: dict | None = None) -> float:
"""Medium tier — stacked Monday conflicts (scenarios/monday_morning.json)."""
return _score_episode(
_as_reward_list(trajectory),
miss_cost=0.09,
overcall_cost=0.04,
stability_gain=0.03,
expertise_gain=0.02,
)
def dinner_disaster_grader(trajectory: dict | None = None) -> float:
"""Hard tier — personal/professional collision (scenarios/dinner_disaster.json)."""
return _score_episode(
_as_reward_list(trajectory),
miss_cost=0.07,
overcall_cost=0.03,
stability_gain=0.02,
expertise_gain=0.04,
)
__all__ = [
"phase2_core_grader",
"monday_morning_grader",
"dinner_disaster_grader",
"STRICT_MIN",
"STRICT_MAX",
]
|