Update inference.py
Browse files- inference.py +14 -12
inference.py
CHANGED
|
@@ -9,15 +9,16 @@ API_BASE_URL = os.environ.get("API_BASE_URL")
|
|
| 9 |
API_KEY = os.environ.get("API_KEY")
|
| 10 |
MODEL_NAME = os.environ.get("MODEL_NAME", "meta-llama/Llama-3-70b-chat-hf")
|
| 11 |
|
|
|
|
| 12 |
client = OpenAI(base_url=API_BASE_URL, api_key=API_KEY)
|
| 13 |
|
| 14 |
def get_llm_action(email_desc, context, keywords):
|
| 15 |
prompt = f"""
|
| 16 |
-
|
|
|
|
| 17 |
Context: {context}
|
| 18 |
Keywords: {keywords}
|
| 19 |
-
|
| 20 |
-
Return only numbers: e.g., 1, 0, 1
|
| 21 |
"""
|
| 22 |
try:
|
| 23 |
response = client.chat.completions.create(
|
|
@@ -29,7 +30,7 @@ def get_llm_action(email_desc, context, keywords):
|
|
| 29 |
res = response.choices[0].message.content.strip()
|
| 30 |
actions = [int(x.strip()) for x in res.split(",")[:3]]
|
| 31 |
return np.array(actions, dtype=np.int64)
|
| 32 |
-
except:
|
| 33 |
return np.array([0, 0, 0], dtype=np.int64)
|
| 34 |
|
| 35 |
def run_inference():
|
|
@@ -37,10 +38,11 @@ def run_inference():
|
|
| 37 |
|
| 38 |
for task_name in tasks:
|
| 39 |
try:
|
| 40 |
-
#
|
| 41 |
env = EmailTriageEnv(task=task_name, shuffle=False)
|
| 42 |
obs, info = env.reset(seed=42)
|
| 43 |
|
|
|
|
| 44 |
print(f"[START] task={task_name}", flush=True)
|
| 45 |
|
| 46 |
terminated = False
|
|
@@ -49,24 +51,24 @@ def run_inference():
|
|
| 49 |
emails = list(env._queue)
|
| 50 |
|
| 51 |
while not terminated and step_count < len(emails):
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
|
|
|
| 55 |
|
| 56 |
obs, reward, terminated, truncated, info = env.step(action)
|
| 57 |
total_reward += reward
|
| 58 |
|
|
|
|
| 59 |
print(f"[STEP] step={step_count+1} reward={reward:.2f}", flush=True)
|
| 60 |
step_count += 1
|
| 61 |
|
| 62 |
-
# --- FIX: SCORE CLIPPING (Strictly between 0 and 1) ---
|
| 63 |
-
# Agar score 1.0 hai toh 0.99 ho jayega, agar 0.0 hai toh 0.01
|
| 64 |
-
final_score = max(0.01, min(0.99, total_reward))
|
| 65 |
|
|
|
|
| 66 |
print(f"[END] task={task_name} score={final_score:.3f} steps={step_count}", flush=True)
|
| 67 |
|
| 68 |
except Exception as e:
|
| 69 |
-
|
| 70 |
print(f"[END] task={task_name} score=0.010 steps=0", flush=True)
|
| 71 |
|
| 72 |
if __name__ == "__main__":
|
|
|
|
| 9 |
API_KEY = os.environ.get("API_KEY")
|
| 10 |
MODEL_NAME = os.environ.get("MODEL_NAME", "meta-llama/Llama-3-70b-chat-hf")
|
| 11 |
|
| 12 |
+
# Proxy Client Initialization
|
| 13 |
client = OpenAI(base_url=API_BASE_URL, api_key=API_KEY)
|
| 14 |
|
| 15 |
def get_llm_action(email_desc, context, keywords):
|
| 16 |
prompt = f"""
|
| 17 |
+
Classify this email:
|
| 18 |
+
Desc: {email_desc}
|
| 19 |
Context: {context}
|
| 20 |
Keywords: {keywords}
|
| 21 |
+
Return 3 numbers (0-2) separated by commas. Example: 1, 0, 1
|
|
|
|
| 22 |
"""
|
| 23 |
try:
|
| 24 |
response = client.chat.completions.create(
|
|
|
|
| 30 |
res = response.choices[0].message.content.strip()
|
| 31 |
actions = [int(x.strip()) for x in res.split(",")[:3]]
|
| 32 |
return np.array(actions, dtype=np.int64)
|
| 33 |
+
except Exception:
|
| 34 |
return np.array([0, 0, 0], dtype=np.int64)
|
| 35 |
|
| 36 |
def run_inference():
|
|
|
|
| 38 |
|
| 39 |
for task_name in tasks:
|
| 40 |
try:
|
| 41 |
+
# Environment initialize
|
| 42 |
env = EmailTriageEnv(task=task_name, shuffle=False)
|
| 43 |
obs, info = env.reset(seed=42)
|
| 44 |
|
| 45 |
+
|
| 46 |
print(f"[START] task={task_name}", flush=True)
|
| 47 |
|
| 48 |
terminated = False
|
|
|
|
| 51 |
emails = list(env._queue)
|
| 52 |
|
| 53 |
while not terminated and step_count < len(emails):
|
| 54 |
+
email_data = emails[step_count]
|
| 55 |
+
action = get_llm_action(email_data['description'],
|
| 56 |
+
email_data['context'],
|
| 57 |
+
email_data['keywords'])
|
| 58 |
|
| 59 |
obs, reward, terminated, truncated, info = env.step(action)
|
| 60 |
total_reward += reward
|
| 61 |
|
| 62 |
+
|
| 63 |
print(f"[STEP] step={step_count+1} reward={reward:.2f}", flush=True)
|
| 64 |
step_count += 1
|
| 65 |
|
|
|
|
|
|
|
|
|
|
| 66 |
|
| 67 |
+
final_score = max(0.01, min(0.99, total_reward))
|
| 68 |
print(f"[END] task={task_name} score={final_score:.3f} steps={step_count}", flush=True)
|
| 69 |
|
| 70 |
except Exception as e:
|
| 71 |
+
|
| 72 |
print(f"[END] task={task_name} score=0.010 steps=0", flush=True)
|
| 73 |
|
| 74 |
if __name__ == "__main__":
|