Spaces:
Sleeping
Sleeping
TanujInsane commited on
Commit ·
9e6cefa
1
Parent(s): c5c9f44
ML agent: TF-IDF cosine similarity + hierarchical partial credit grading
Browse files- grading.py +149 -257
grading.py
CHANGED
|
@@ -1,284 +1,176 @@
|
|
| 1 |
"""
|
| 2 |
-
Grade agents on
|
| 3 |
-
|
| 4 |
"""
|
| 5 |
|
| 6 |
import numpy as np
|
| 7 |
-
from typing import Dict, List
|
| 8 |
from environment import DocumentClassificationEnv
|
| 9 |
-
|
| 10 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 11 |
|
| 12 |
|
| 13 |
class AgentGrader:
|
| 14 |
-
"""
|
| 15 |
-
|
| 16 |
-
def __init__(self, task_difficulty: str):
|
| 17 |
-
"""Set up the grader for a specific difficulty"""
|
| 18 |
self.task_difficulty = task_difficulty
|
|
|
|
| 19 |
self.env = DocumentClassificationEnv(task_difficulty=task_difficulty, seed=42)
|
| 20 |
-
|
| 21 |
-
def
|
| 22 |
"""
|
| 23 |
-
Run agent through
|
| 24 |
-
|
| 25 |
Args:
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
Returns:
|
| 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 |
-
predictions,
|
| 71 |
-
ground_truths,
|
| 72 |
-
rewards,
|
| 73 |
-
processing_times,
|
| 74 |
-
total_time
|
| 75 |
-
)
|
| 76 |
-
|
| 77 |
-
# Get final score
|
| 78 |
-
score = self._calculate_score(metrics)
|
| 79 |
-
|
| 80 |
-
if verbose:
|
| 81 |
-
self._print_results(metrics, score)
|
| 82 |
-
|
| 83 |
-
return score, metrics
|
| 84 |
-
|
| 85 |
-
def _calculate_metrics(self,
|
| 86 |
-
predictions: List[int],
|
| 87 |
-
ground_truths: List[str],
|
| 88 |
-
rewards: List[float],
|
| 89 |
-
processing_times: List[float],
|
| 90 |
-
total_time: float) -> Dict:
|
| 91 |
-
"""Calculate performance metrics"""
|
| 92 |
-
|
| 93 |
-
# Convert ground truth category names to indices
|
| 94 |
-
env_categories = list(self.env.CATEGORY_MAPS[self.task_difficulty].values())
|
| 95 |
-
ground_truth_indices = []
|
| 96 |
-
for gt in ground_truths:
|
| 97 |
-
try:
|
| 98 |
-
idx = env_categories.index(gt)
|
| 99 |
-
ground_truth_indices.append(idx)
|
| 100 |
-
except ValueError:
|
| 101 |
-
continue
|
| 102 |
-
|
| 103 |
-
predictions = predictions[:len(ground_truth_indices)]
|
| 104 |
-
|
| 105 |
-
accuracy = np.mean(np.array(predictions) == np.array(ground_truth_indices)) if len(predictions) > 0 else 0.0
|
| 106 |
-
|
| 107 |
metrics = {
|
| 108 |
-
"accuracy":
|
| 109 |
-
"
|
| 110 |
-
"
|
| 111 |
-
"
|
| 112 |
-
"total_reward": np.sum(rewards) if rewards else 0.0,
|
| 113 |
-
"average_processing_time_ms": np.mean(processing_times) * 1000 if processing_times else 0.0,
|
| 114 |
-
"total_time_seconds": total_time,
|
| 115 |
}
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
|
| 126 |
-
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
ground_truth_indices, predictions,
|
| 130 |
-
average='macro', zero_division=0
|
| 131 |
-
)
|
| 132 |
-
except:
|
| 133 |
-
pass
|
| 134 |
-
|
| 135 |
-
return metrics
|
| 136 |
-
|
| 137 |
-
def _calculate_score(self, metrics: Dict) -> float:
|
| 138 |
-
"""
|
| 139 |
-
Calculate final score.
|
| 140 |
-
Easy = just accuracy
|
| 141 |
-
Medium = 80% accuracy + 20% speed bonus
|
| 142 |
-
Hard = 75% accuracy + 25% speed bonus
|
| 143 |
-
"""
|
| 144 |
accuracy = metrics["accuracy"]
|
| 145 |
-
|
| 146 |
-
|
|
|
|
| 147 |
if self.task_difficulty == "easy":
|
| 148 |
-
|
| 149 |
-
|
| 150 |
-
|
|
|
|
| 151 |
elif self.task_difficulty == "medium":
|
| 152 |
-
# Medium: care about both
|
| 153 |
time_bonus = 0.0
|
| 154 |
-
avg_time = metrics["average_processing_time_ms"]
|
| 155 |
if avg_time < 200:
|
| 156 |
time_bonus = 0.15
|
| 157 |
elif avg_time < 500:
|
| 158 |
-
time_bonus = 0.
|
| 159 |
-
|
| 160 |
-
score = 0.
|
| 161 |
-
|
| 162 |
-
else: # hard
|
| 163 |
-
# Hard: need speed to compensate for lower accuracy
|
| 164 |
time_bonus = 0.0
|
| 165 |
-
avg_time = metrics["average_processing_time_ms"]
|
| 166 |
if avg_time < 100:
|
| 167 |
time_bonus = 0.25
|
| 168 |
elif avg_time < 300:
|
| 169 |
-
time_bonus = 0.
|
| 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 |
-
print(f"{
|
| 197 |
-
|
| 198 |
-
|
| 199 |
-
class BaselineAgent:
|
| 200 |
-
"""Simple baseline - just uses keyword matching"""
|
| 201 |
-
|
| 202 |
-
def __init__(self, difficulty: str):
|
| 203 |
-
self.difficulty = difficulty
|
| 204 |
-
|
| 205 |
-
def decide(self, observation):
|
| 206 |
-
"""Look for keywords in the document, make a guess"""
|
| 207 |
-
content = observation.get("content", "").lower()
|
| 208 |
-
features = observation.get("features", [])
|
| 209 |
-
|
| 210 |
-
if self.difficulty == "easy":
|
| 211 |
-
# Easy mode - just look for obvious keywords
|
| 212 |
-
if "billing" in content or "bill" in content or "invoice" in content:
|
| 213 |
-
return 1 # Billing
|
| 214 |
-
elif "support" in content or "help" in content or "issue" in content:
|
| 215 |
-
return 2 # Support
|
| 216 |
-
elif "technical" in content or "bug" in content or "error" in content:
|
| 217 |
-
return 3 # Technical
|
| 218 |
-
elif "hr" in content or "payroll" in content or "benefits" in content:
|
| 219 |
-
return 4 # HR
|
| 220 |
-
return 0 # Default: General
|
| 221 |
-
|
| 222 |
-
elif self.difficulty == "medium":
|
| 223 |
-
# Medium - more keywords to look for
|
| 224 |
-
if "dispute" in content or "overcharg" in content:
|
| 225 |
-
return 2 # Billing-Dispute
|
| 226 |
-
elif "urgent" in content or "critical" in content or "emergency" in content:
|
| 227 |
-
return 3 # Support (using urgent as proxy)
|
| 228 |
-
elif "bug" in content or "error" in content or "crash" in content:
|
| 229 |
-
return 5 # Technical-Bug
|
| 230 |
-
elif "payroll" in content or "salary" in content or "payment" in content:
|
| 231 |
-
return 6 # HR-Payroll
|
| 232 |
-
elif "benefits" in content or "health" in content or "enroll" in content:
|
| 233 |
-
return 7 # HR-Benefits
|
| 234 |
-
elif "billing" in content or "invoice" in content:
|
| 235 |
-
return 1 # Billing
|
| 236 |
-
elif "technical" in content:
|
| 237 |
-
return 4 # Technical
|
| 238 |
-
elif "legal" in content or "contract" in content:
|
| 239 |
-
return 8 # Legal
|
| 240 |
-
elif "executive" in content or "manager" in content or "strategy" in content:
|
| 241 |
-
return 9 # Executive
|
| 242 |
-
return 0 # General
|
| 243 |
-
|
| 244 |
-
else: # hard
|
| 245 |
-
# Hard mode - need more keywords and combinations
|
| 246 |
-
if "urgent" in content or "critical" in content:
|
| 247 |
-
return 4 # Support-Urgent
|
| 248 |
-
elif "dispute" in content or "overcharge" in content:
|
| 249 |
-
return 2 # Billing-Dispute
|
| 250 |
-
elif "refund" in content:
|
| 251 |
-
return 3 # Billing-Refund
|
| 252 |
-
elif "bug" in content or "crash" in content:
|
| 253 |
-
return 7 # Technical-Bug
|
| 254 |
-
elif "feature" in content or "request" in content:
|
| 255 |
-
return 8 # Technical-Feature
|
| 256 |
-
elif "payroll" in content:
|
| 257 |
-
return 9 # HR-Payroll
|
| 258 |
-
elif "benefits" in content:
|
| 259 |
-
return 10 # HR-Benefits
|
| 260 |
-
elif "complaint" in content or "complain" in content:
|
| 261 |
-
return 11 # HR-Complaint
|
| 262 |
-
elif "contract" in content:
|
| 263 |
-
return 13 # Legal-Contract
|
| 264 |
-
elif "compliance" in content:
|
| 265 |
-
return 14 # Legal-Compliance
|
| 266 |
-
elif "strategic" in content or "strategy" in content:
|
| 267 |
-
return 16 # Executive-Strategic
|
| 268 |
-
elif "billing" in content or "invoice" in content:
|
| 269 |
-
return 1 # Billing
|
| 270 |
-
elif "support" in content:
|
| 271 |
-
return 5 # Support-Normal
|
| 272 |
-
elif "technical" in content:
|
| 273 |
-
return 6 # Technical
|
| 274 |
-
elif "legal" in content:
|
| 275 |
-
return 12 # Legal
|
| 276 |
-
elif "finance" in content:
|
| 277 |
-
return 17 # Finance
|
| 278 |
-
elif "marketing" in content:
|
| 279 |
-
return 18 # Marketing
|
| 280 |
-
elif "operations" in content:
|
| 281 |
-
return 19 # Operations
|
| 282 |
-
elif "executive" in content:
|
| 283 |
-
return 15 # Executive
|
| 284 |
-
return 0 # General
|
|
|
|
| 1 |
"""
|
| 2 |
+
Grade agents on document classification.
|
| 3 |
+
Uses hierarchical partial credit - related categories get partial score.
|
| 4 |
"""
|
| 5 |
|
| 6 |
import numpy as np
|
| 7 |
+
from typing import Dict, List
|
| 8 |
from environment import DocumentClassificationEnv
|
| 9 |
+
|
| 10 |
+
|
| 11 |
+
# Hierarchy map - categories that are "close" get partial credit
|
| 12 |
+
CATEGORY_HIERARCHY = {
|
| 13 |
+
"Billing": ["Billing-Dispute", "Billing-Refund"],
|
| 14 |
+
"Billing-Dispute": ["Billing", "Billing-Refund"],
|
| 15 |
+
"Billing-Refund": ["Billing", "Billing-Dispute"],
|
| 16 |
+
"Support": ["Support-Urgent", "Support-Normal"],
|
| 17 |
+
"Support-Urgent": ["Support", "Support-Normal"],
|
| 18 |
+
"Support-Normal": ["Support", "Support-Urgent"],
|
| 19 |
+
"Technical": ["Technical-Bug", "Technical-Feature"],
|
| 20 |
+
"Technical-Bug": ["Technical", "Technical-Feature"],
|
| 21 |
+
"Technical-Feature": ["Technical", "Technical-Bug"],
|
| 22 |
+
"HR": ["HR-Payroll", "HR-Benefits", "HR-Complaint"],
|
| 23 |
+
"HR-Payroll": ["HR", "HR-Benefits"],
|
| 24 |
+
"HR-Benefits": ["HR", "HR-Payroll"],
|
| 25 |
+
"HR-Complaint": ["HR"],
|
| 26 |
+
"Legal": ["Legal-Contract", "Legal-Compliance"],
|
| 27 |
+
"Legal-Contract": ["Legal", "Legal-Compliance"],
|
| 28 |
+
"Legal-Compliance": ["Legal", "Legal-Contract"],
|
| 29 |
+
"Executive": ["Executive-Strategic"],
|
| 30 |
+
"Executive-Strategic": ["Executive"],
|
| 31 |
+
}
|
| 32 |
+
|
| 33 |
+
PARTIAL_CREDIT = 0.4 # Score for picking a related category
|
| 34 |
+
|
| 35 |
+
|
| 36 |
+
def hierarchical_accuracy(true_label: str, pred_label: str) -> float:
|
| 37 |
+
"""
|
| 38 |
+
Returns:
|
| 39 |
+
1.0 if exact match
|
| 40 |
+
0.4 if same parent category (e.g. Billing vs Billing-Dispute)
|
| 41 |
+
0.0 otherwise
|
| 42 |
+
"""
|
| 43 |
+
if true_label == pred_label:
|
| 44 |
+
return 1.0
|
| 45 |
+
related = CATEGORY_HIERARCHY.get(true_label, [])
|
| 46 |
+
if pred_label in related:
|
| 47 |
+
return PARTIAL_CREDIT
|
| 48 |
+
return 0.0
|
| 49 |
|
| 50 |
|
| 51 |
class AgentGrader:
|
| 52 |
+
"""Grade an agent on a specific task difficulty"""
|
| 53 |
+
|
| 54 |
+
def __init__(self, task_difficulty: str, num_episodes: int = 3):
|
|
|
|
| 55 |
self.task_difficulty = task_difficulty
|
| 56 |
+
self.num_episodes = num_episodes
|
| 57 |
self.env = DocumentClassificationEnv(task_difficulty=task_difficulty, seed=42)
|
| 58 |
+
|
| 59 |
+
def grade(self, agent_fn) -> Dict:
|
| 60 |
"""
|
| 61 |
+
Run agent through multiple episodes and return score 0.0-1.0.
|
| 62 |
+
|
| 63 |
Args:
|
| 64 |
+
agent_fn: callable(observation) -> action (int)
|
| 65 |
+
|
|
|
|
| 66 |
Returns:
|
| 67 |
+
dict with score and detailed metrics
|
| 68 |
"""
|
| 69 |
+
all_scores = []
|
| 70 |
+
all_accuracies = []
|
| 71 |
+
all_partial_credits = []
|
| 72 |
+
|
| 73 |
+
for ep in range(self.num_episodes):
|
| 74 |
+
obs, _ = self.env.reset(seed=ep * 777)
|
| 75 |
+
ep_true = []
|
| 76 |
+
ep_pred = []
|
| 77 |
+
ep_rewards = []
|
| 78 |
+
terminated = False
|
| 79 |
+
|
| 80 |
+
while not terminated:
|
| 81 |
+
action = agent_fn(obs)
|
| 82 |
+
obs, reward, terminated, _, info = self.env.step(action)
|
| 83 |
+
ep_true.append(info.get("true_category", ""))
|
| 84 |
+
pred_cat = info.get("predicted_category", "")
|
| 85 |
+
ep_pred.append(pred_cat)
|
| 86 |
+
ep_rewards.append(reward)
|
| 87 |
+
|
| 88 |
+
# Exact accuracy
|
| 89 |
+
exact = sum(t == p for t, p in zip(ep_true, ep_pred)) / len(ep_true)
|
| 90 |
+
|
| 91 |
+
# Hierarchical accuracy (with partial credit)
|
| 92 |
+
hier_scores = [hierarchical_accuracy(t, p) for t, p in zip(ep_true, ep_pred)]
|
| 93 |
+
hier_acc = np.mean(hier_scores)
|
| 94 |
+
|
| 95 |
+
# Partial credit ratio
|
| 96 |
+
partial = sum(1 for s in hier_scores if 0 < s < 1) / len(hier_scores)
|
| 97 |
+
|
| 98 |
+
all_accuracies.append(exact)
|
| 99 |
+
all_scores.append(hier_acc)
|
| 100 |
+
all_partial_credits.append(partial)
|
| 101 |
+
|
| 102 |
+
avg_accuracy = np.mean(all_accuracies)
|
| 103 |
+
avg_hier = np.mean(all_scores)
|
| 104 |
+
avg_partial = np.mean(all_partial_credits)
|
| 105 |
+
|
| 106 |
+
# Compute final score based on difficulty
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 107 |
metrics = {
|
| 108 |
+
"accuracy": avg_accuracy,
|
| 109 |
+
"hierarchical_accuracy": avg_hier,
|
| 110 |
+
"partial_credit_rate": avg_partial,
|
| 111 |
+
"average_processing_time_ms": 50, # placeholder
|
|
|
|
|
|
|
|
|
|
| 112 |
}
|
| 113 |
+
|
| 114 |
+
score = self._compute_score(metrics)
|
| 115 |
+
|
| 116 |
+
return {
|
| 117 |
+
"score": score,
|
| 118 |
+
"accuracy": avg_accuracy,
|
| 119 |
+
"hierarchical_accuracy": avg_hier,
|
| 120 |
+
"partial_credit_rate": avg_partial,
|
| 121 |
+
"task_difficulty": self.task_difficulty,
|
| 122 |
+
"num_episodes": self.num_episodes,
|
| 123 |
+
}
|
| 124 |
+
|
| 125 |
+
def _compute_score(self, metrics: Dict) -> float:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 126 |
accuracy = metrics["accuracy"]
|
| 127 |
+
hier_accuracy = metrics["hierarchical_accuracy"]
|
| 128 |
+
avg_time = metrics.get("average_processing_time_ms", 100)
|
| 129 |
+
|
| 130 |
if self.task_difficulty == "easy":
|
| 131 |
+
time_bonus = 0.1 if avg_time < 100 else 0.0
|
| 132 |
+
# Easy: mostly exact accuracy
|
| 133 |
+
score = 0.85 * accuracy + 0.05 * hier_accuracy + 0.1 * time_bonus
|
| 134 |
+
|
| 135 |
elif self.task_difficulty == "medium":
|
|
|
|
| 136 |
time_bonus = 0.0
|
|
|
|
| 137 |
if avg_time < 200:
|
| 138 |
time_bonus = 0.15
|
| 139 |
elif avg_time < 500:
|
| 140 |
+
time_bonus = 0.10
|
| 141 |
+
# Medium: mix of exact + hierarchical
|
| 142 |
+
score = 0.70 * accuracy + 0.15 * hier_accuracy + 0.15 * time_bonus
|
| 143 |
+
|
| 144 |
+
else: # hard / extreme
|
|
|
|
| 145 |
time_bonus = 0.0
|
|
|
|
| 146 |
if avg_time < 100:
|
| 147 |
time_bonus = 0.25
|
| 148 |
elif avg_time < 300:
|
| 149 |
+
time_bonus = 0.15
|
| 150 |
+
elif avg_time < 500:
|
| 151 |
+
time_bonus = 0.05
|
| 152 |
+
# Hard: hierarchical accuracy matters more (22+ categories)
|
| 153 |
+
score = 0.50 * accuracy + 0.30 * hier_accuracy + 0.20 * time_bonus
|
| 154 |
+
|
| 155 |
+
return min(1.0, max(0.0, score))
|
| 156 |
+
|
| 157 |
+
|
| 158 |
+
def grade_baseline(task_difficulty: str) -> float:
|
| 159 |
+
"""Quick helper to grade and return score"""
|
| 160 |
+
from baseline_inference import TFIDFAgent
|
| 161 |
+
|
| 162 |
+
agent = TFIDFAgent(task_difficulty)
|
| 163 |
+
grader = AgentGrader(task_difficulty)
|
| 164 |
+
|
| 165 |
+
def agent_fn(obs):
|
| 166 |
+
return agent.predict(obs["content"])
|
| 167 |
+
|
| 168 |
+
result = grader.grade(agent_fn)
|
| 169 |
+
return result["score"]
|
| 170 |
+
|
| 171 |
+
|
| 172 |
+
if __name__ == "__main__":
|
| 173 |
+
print("Grading baseline agent with hierarchical partial credit...\n")
|
| 174 |
+
for difficulty in ["easy", "medium", "hard"]:
|
| 175 |
+
score = grade_baseline(difficulty)
|
| 176 |
+
print(f" {difficulty.upper()} Score: {score:.4f}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|