Spaces:
Sleeping
Sleeping
File size: 7,047 Bytes
18feac5 | 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 | # from tool_use_env.client import ToolUseEnv
# from tool_use_env.models import ToolUseAction
# import random
# def rule_based_policy(query: str):
# query = query.lower()
# # --- Introduce slight imperfection ---
# if random.random() < 0.1:
# return "answer_directly"
# if "what is" in query and any(op in query for op in ["+", "-", "*", "/"]):
# return "use_calculator"
# if "capital" in query or "who is" in query:
# return "use_search"
# return "answer_directly"
# def run_single_episode(env):
# result = env.reset()
# obs = result.observation
# query = obs.query
# action_type = rule_based_policy(query)
# action = ToolUseAction(action_type=action_type)
# result = env.step(action)
# obs = result.observation
# return {
# "query": query,
# "action": action_type,
# "reward": result.reward,
# "message": obs.message
# }
# def run_evaluation(num_episodes=20):
# results = []
# difficulty_scores = {
# "easy": [],
# "medium": [],
# "hard": []
# }
# total_score = 0
# with ToolUseEnv(base_url="http://localhost:8000").sync() as env:
# for _ in range(num_episodes):
# result = env.reset()
# obs = result.observation
# query = obs.query
# state = env.state()
# difficulty = state.difficulty
# action_type = rule_based_policy(query)
# action = ToolUseAction(action_type=action_type)
# result = env.step(action)
# score = result.reward
# total_score += score
# difficulty_scores[difficulty].append(score)
# results.append({
# "query": query,
# "difficulty": difficulty,
# "action": action_type,
# "score": score,
# "message": result.observation.message
# })
# avg_score = total_score / num_episodes
# print("\n=== OVERALL PERFORMANCE ===")
# print(f"Average Score: {avg_score:.2f}")
# print("\n=== DIFFICULTY BREAKDOWN ===")
# for level in difficulty_scores:
# if difficulty_scores[level]:
# avg = sum(difficulty_scores[level]) / len(difficulty_scores[level])
# print(f"{level.capitalize()}: {avg:.2f}")
# print("\n=== SAMPLE CASES ===")
# for r in results[:5]:
# print(f"\nQuery: {r['query']}")
# print(f"Action: {r['action']}")
# print(f"Score: {r['score']:.2f}")
# print(f"Details: {r['message']}")
# return results
# def analyze_failures(results):
# wrong_decisions = 0
# tool_failures = 0
# total = len(results)
# for r in results:
# msg = r["message"]
# if "Correct: False" in msg:
# if "use_" in msg:
# tool_failures += 1
# else:
# wrong_decisions += 1
# print("\n=== FAILURE ANALYSIS ===")
# print(f"Tool failures: {tool_failures}/{total} ({(tool_failures/total)*100:.1f}%)")
# print(f"Wrong decisions: {wrong_decisions}/{total} ({(wrong_decisions/total)*100:.1f}%)")
# if __name__ == "__main__":
# results = run_evaluation(50)
# analyze_failures(results)
import os
import random
from collections import defaultdict
from dotenv import load_dotenv
from openai import OpenAI
from tool_use_env.client import ToolUseEnv
from tool_use_env.models import ToolUseAction
# --- Load environment variables ---
load_dotenv()
# --- Initialize OpenAI client ---
client = OpenAI()
# --- Reproducibility ---
random.seed(42)
# 🧠 LLM Policy (CORE)
def llm_policy(query: str):
prompt = f"""
You are an AI agent choosing the best tool.
Available actions:
- use_calculator (for math problems)
- use_search (for factual questions)
- answer_directly (if neither tool is needed)
Query: {query}
Respond with ONLY one of:
use_calculator
use_search
answer_directly
"""
try:
response = client.chat.completions.create(
model="gpt-4o-mini",
messages=[{"role": "user", "content": prompt}],
temperature=0
)
action = response.choices[0].message.content.strip()
# --- Safety check ---
if action not in ["use_calculator", "use_search", "answer_directly"]:
return "answer_directly"
return action
except Exception as e:
print(f"[ERROR] LLM call failed: {e}")
return "answer_directly"
# 🧪 Evaluation Loop
def run_evaluation(num_episodes=50):
results = []
total_score = 0
difficulty_scores = defaultdict(list)
with ToolUseEnv(base_url="http://localhost:8000").sync() as env:
for _ in range(num_episodes):
# --- Reset ---
result = env.reset()
obs = result.observation
query = obs.query
# --- Get difficulty ---
state = env.state()
difficulty = state.difficulty
# --- LLM decides action ---
action_type = llm_policy(query)
action = ToolUseAction(action_type=action_type)
# --- Step ---
result = env.step(action)
obs = result.observation
score = result.reward
total_score += score
difficulty_scores[difficulty].append(score)
results.append({
"query": query,
"difficulty": difficulty,
"action": action_type,
"score": score,
"message": obs.message
})
print(f"Score: {score:.2f}")
# --- Overall ---
avg_score = total_score / num_episodes
print("\n=== OVERALL PERFORMANCE ===")
print(f"Average Score: {avg_score:.2f}")
# --- Breakdown ---
print("\n=== DIFFICULTY BREAKDOWN ===")
for level in ["easy", "medium", "hard"]:
if difficulty_scores[level]:
avg = sum(difficulty_scores[level]) / len(difficulty_scores[level])
print(f"{level.capitalize()}: {avg:.2f}")
# --- Sample Cases ---
print("\n=== SAMPLE CASES ===")
for r in results[:5]:
print(f"\nQuery: {r['query']}")
print(f"Action: {r['action']}")
print(f"Score: {r['score']:.2f}")
print(f"Details: {r['message']}")
return results
# 📊 Failure Analysis
def analyze_failures(results):
total = len(results)
tool_failures = 0
wrong_decisions = 0
for r in results:
msg = r["message"]
if "Correct: False" in msg:
if "use_" in msg:
tool_failures += 1
else:
wrong_decisions += 1
print("\n=== FAILURE ANALYSIS ===")
print(f"Tool failures: {tool_failures}/{total} ({(tool_failures/total)*100:.1f}%)")
print(f"Wrong decisions: {wrong_decisions}/{total} ({(wrong_decisions/total)*100:.1f}%)")
# 🚀 Main
if __name__ == "__main__":
results = run_evaluation(50)
analyze_failures(results) |