TanujInsane
feat: richer grader output object
9260b23
Raw History Blame Contribute Delete
9.5 kB
"""
Grade agents on document classification.
Uses hierarchical partial credit - related categories get partial score.
Handles tool actions (metadata requests, escalation) gracefully.
"""
import numpy as np
import json
from typing import Dict, List, Callable
from environment import DocumentClassificationEnv
from constants import CATEGORY_HIERARCHY
# Hierarchy map - categories that are "close" get partial credit
# CATEGORY_HIERARCHY is now in constants.py
PARTIAL_CREDIT = 0.4 # Score for picking a related category
TASK_ID_MAP = {
"easy": 1,
"medium": 2,
"hard": 3,
}
def hierarchical_accuracy(true_label: str, pred_label: str) -> float:
"""
Returns:
1.0 if exact match
0.4 if same parent category (e.g. Billing vs Billing-Dispute)
0.0 otherwise (including TOOL_METADATA and ESCALATED)
"""
if pred_label in ("TOOL_METADATA", "ESCALATED"):
return 0.0
if true_label == pred_label:
return 1.0
related = CATEGORY_HIERARCHY.get(true_label, [])
if pred_label in related:
return PARTIAL_CREDIT
return 0.0
class AgentGrader:
"""Grade an agent on a specific task difficulty.
Handles the full action space including tool actions:
- Standard classification actions (0..N-1)
- Request Metadata tool (action N)
- Escalate to Human (action N+1)
"""
def __init__(self, task_difficulty: str, num_episodes: int = 3):
self.task_difficulty = task_difficulty
self.num_episodes = num_episodes
self.env = DocumentClassificationEnv(task_difficulty=task_difficulty, seed=42)
def grade(self, agent_fn: Callable) -> Dict:
"""
Run agent through multiple episodes and return score 0.0-1.0.
Args:
agent_fn: callable(observation) -> action (int)
Returns:
dict with score and detailed metrics
"""
all_scores = []
all_accuracies = []
all_partial_credits = []
all_tool_usage_rates = []
all_escalation_rates = []
all_agent_times = []
all_ep_true = []
all_ep_pred = []
import time
for ep in range(self.num_episodes):
obs, _ = self.env.reset(seed=ep * 777)
ep_true = []
ep_pred = []
ep_rewards = []
tool_requests = 0
escalations = 0
classifications = 0
terminated = False
while not terminated:
t0 = time.time()
action = agent_fn(obs)
t1 = time.time()
all_agent_times.append((t1 - t0) * 1000)
obs, reward, terminated, _, info = self.env.step(action)
pred_cat = info.get("predicted_category", "")
true_cat = info.get("true_category", "")
if pred_cat == "TOOL_METADATA":
tool_requests += 1
# Don't count tool usage as a classification attempt
continue
elif pred_cat == "ESCALATED":
escalations += 1
ep_true.append(true_cat)
ep_pred.append(pred_cat)
ep_rewards.append(reward)
classifications += 1
else:
ep_true.append(true_cat)
ep_pred.append(pred_cat)
ep_rewards.append(reward)
classifications += 1
if len(ep_true) == 0:
all_accuracies.append(0.0)
all_scores.append(0.0)
all_partial_credits.append(0.0)
all_tool_usage_rates.append(0.0)
all_escalation_rates.append(0.0)
continue
# Exact accuracy (only count real classifications, not escalations)
real_classifications = [(t, p) for t, p in zip(ep_true, ep_pred) if p != "ESCALATED"]
if real_classifications:
exact = sum(t == p for t, p in real_classifications) / len(real_classifications)
else:
exact = 0.0
# Hierarchical accuracy (with partial credit)
hier_scores = [
hierarchical_accuracy(t, p) for t, p in zip(ep_true, ep_pred)
]
hier_acc = np.mean(hier_scores) if hier_scores else 0.0
# Partial credit ratio
partial = sum(1 for s in hier_scores if 0 < s < 1) / len(hier_scores) if hier_scores else 0.0
# Tool usage metrics
total_actions = classifications + tool_requests
tool_rate = tool_requests / total_actions if total_actions > 0 else 0.0
esc_rate = escalations / classifications if classifications > 0 else 0.0
all_accuracies.append(exact)
all_scores.append(hier_acc)
all_partial_credits.append(partial)
all_tool_usage_rates.append(tool_rate)
all_escalation_rates.append(esc_rate)
all_ep_true.append(ep_true)
all_ep_pred.append(ep_pred)
avg_accuracy = float(np.mean(all_accuracies))
avg_hier = float(np.mean(all_scores))
avg_partial = float(np.mean(all_partial_credits))
avg_tool_rate = float(np.mean(all_tool_usage_rates))
avg_esc_rate = float(np.mean(all_escalation_rates))
avg_processing_time = float(np.mean(all_agent_times)) if all_agent_times else 100.0
# Generate Confusion Matrix
from collections import defaultdict
cm = defaultdict(lambda: defaultdict(int))
for ep_t, ep_p in zip(all_ep_true, all_ep_pred):
for t_label, p_label in zip(ep_t, ep_p):
cm[t_label][p_label] += 1
cm_dict = {k: dict(v) for k, v in cm.items()}
try:
out_path = f"/tmp/confusion_matrix_{self.task_difficulty}.json"
with open(out_path, "w") as f:
json.dump(cm_dict, f, indent=2)
except Exception:
pass
# Compute final score based on difficulty
metrics = {
"accuracy": avg_accuracy,
"hierarchical_accuracy": avg_hier,
"partial_credit_rate": avg_partial,
"average_processing_time_ms": avg_processing_time,
"tool_usage_rate": avg_tool_rate,
"escalation_rate": avg_esc_rate,
}
score_raw = self._compute_score(metrics)
score_clamped = float(min(1.0, max(0.0, score_raw)))
import uuid
task_id = TASK_ID_MAP.get(self.task_difficulty, 0)
episode_id = str(uuid.uuid4())
scenario_id = f"document_classification_{self.task_difficulty}"
breakdown = {
"ACCURACY": avg_accuracy,
"HIERARCHICAL_ACCURACY": avg_hier,
"PARTIAL_CREDIT_RATE": avg_partial,
"AVERAGE_PROCESSING_TIME_MS": avg_processing_time,
"TOOL_USAGE_RATE": avg_tool_rate,
"ESCALATION_RATE": avg_esc_rate,
}
return {
"score": score_clamped,
"task_id": task_id,
"episode_id": episode_id,
"scenario_id": scenario_id,
"breakdown": breakdown,
"grader_version": "1.0.0",
"accuracy": avg_accuracy,
"hierarchical_accuracy": avg_hier,
"partial_credit_rate": avg_partial,
"tool_usage_rate": avg_tool_rate,
"escalation_rate": avg_esc_rate,
"task_difficulty": self.task_difficulty,
"num_episodes": self.num_episodes,
}
def _compute_score(self, metrics: Dict) -> float:
accuracy = metrics["accuracy"]
hier_accuracy = metrics["hierarchical_accuracy"]
avg_time = metrics.get("average_processing_time_ms", 100)
if self.task_difficulty == "easy":
time_bonus = 0.1 if avg_time < 100 else 0.0
# Easy: mostly exact accuracy
score = 0.85 * accuracy + 0.05 * hier_accuracy + 0.1 * time_bonus
elif self.task_difficulty == "medium":
time_bonus = 0.0
if avg_time < 200:
time_bonus = 0.15
elif avg_time < 500:
time_bonus = 0.10
# Medium: mix of exact + hierarchical
score = 0.70 * accuracy + 0.15 * hier_accuracy + 0.15 * time_bonus
else: # hard
time_bonus = 0.0
if avg_time < 100:
time_bonus = 0.25
elif avg_time < 300:
time_bonus = 0.15
elif avg_time < 500:
time_bonus = 0.05
# Hard: hierarchical accuracy matters more (22 categories)
score = 0.50 * accuracy + 0.30 * hier_accuracy + 0.20 * time_bonus
return min(1.0, max(0.0, score))
def grade_baseline(task_difficulty: str) -> float:
"""Quick helper to grade the TF-IDF baseline and return score."""
from baseline_inference import load_or_train
model = load_or_train(task_difficulty)
grader = AgentGrader(task_difficulty)
def agent_fn(obs):
pred = model.predict([obs["content"]])[0]
return int(pred)
result = grader.grade(agent_fn)
return result["score"]
if __name__ == "__main__":
print("Grading baseline agent with hierarchical partial credit...\n")
for difficulty in ["easy", "medium", "hard"]:
score = grade_baseline(difficulty)
print(f" {difficulty.upper()} Score: {score:.4f}")