Muditsharmaaa commited on
Commit
72ad6e3
·
1 Parent(s): 84922be

Fix Output Parsing: clamped terminal logs and valid JSON

Browse files
Files changed (1) hide show
  1. inference.py +14 -15
inference.py CHANGED
@@ -9,10 +9,11 @@ from env import BlockchainEnv
9
 
10
  # --- Mandatory Hackathon Variables ---
11
  API_BASE_URL = os.getenv("API_BASE_URL", "https://api.openai.com/v1")
12
- MODEL_NAME = os.getenv("MODEL_NAME", "gpt-4o-mini") # Swap to your active model
13
  HF_TOKEN = os.getenv("HF_TOKEN")
14
  API_KEY = HF_TOKEN or os.getenv("OPENAI_API_KEY")
15
- TASK_NAME = os.getenv("MY_ENV_TASK", "hard_gas_management")
 
16
  BENCHMARK = "blockchain_certificate_admin"
17
  MAX_STEPS = 5
18
 
@@ -58,7 +59,6 @@ def main():
58
 
59
  log_start(task=TASK_NAME, env=BENCHMARK, model=MODEL_NAME)
60
 
61
- # Initialize Environment
62
  result = env.reset()
63
 
64
  try:
@@ -70,33 +70,33 @@ def main():
70
  user_prompt = f"Pending Requests: {obs.pending_requests}\nGas Balance: {obs.gas_balance}\nStep: {step}"
71
 
72
  try:
73
- # Call LLM
74
  completion = client.chat.completions.create(
75
  model=MODEL_NAME,
76
  messages=[
77
  {"role": "system", "content": SYSTEM_PROMPT},
78
  {"role": "user", "content": user_prompt},
79
  ],
80
- temperature=0.0, # Deterministic logic preferred here
81
  )
82
  raw_response = (completion.choices[0].message.content or "").strip()
83
 
84
- # Parse JSON and execute action
85
  action_dict = json.loads(raw_response)
86
  action = BlockchainAction(**action_dict)
87
  result = env.step(action)
88
 
89
  reward = result.reward
90
  error = None
91
- action_str = json.dumps(action_dict).replace(" ", "") # Compress for STDOUT
92
 
93
  except Exception as e:
94
- # If the LLM hallucinates or JSON parsing fails
95
- reward = -1.0
96
- error = str(e).replace("\n", " ")
97
- action_str = "invalid_format"
98
- result.done = True # End episode on fatal format error
99
 
 
 
100
  rewards.append(reward)
101
  steps_taken = step
102
 
@@ -105,9 +105,8 @@ def main():
105
  if result.done:
106
  break
107
 
108
- # Calculate final normalized score
109
- score = sum(rewards) / steps_taken if steps_taken > 0 else 0.0
110
- score = max(0.0, min(1.0, score))
111
  success = score >= 0.6
112
 
113
  finally:
 
9
 
10
  # --- Mandatory Hackathon Variables ---
11
  API_BASE_URL = os.getenv("API_BASE_URL", "https://api.openai.com/v1")
12
+ MODEL_NAME = os.getenv("MODEL_NAME", "gpt-4o-mini")
13
  HF_TOKEN = os.getenv("HF_TOKEN")
14
  API_KEY = HF_TOKEN or os.getenv("OPENAI_API_KEY")
15
+ # FIX 1: Match the exact ID from your openenv.yaml
16
+ TASK_NAME = os.getenv("MY_ENV_TASK", "easy_minting")
17
  BENCHMARK = "blockchain_certificate_admin"
18
  MAX_STEPS = 5
19
 
 
59
 
60
  log_start(task=TASK_NAME, env=BENCHMARK, model=MODEL_NAME)
61
 
 
62
  result = env.reset()
63
 
64
  try:
 
70
  user_prompt = f"Pending Requests: {obs.pending_requests}\nGas Balance: {obs.gas_balance}\nStep: {step}"
71
 
72
  try:
 
73
  completion = client.chat.completions.create(
74
  model=MODEL_NAME,
75
  messages=[
76
  {"role": "system", "content": SYSTEM_PROMPT},
77
  {"role": "user", "content": user_prompt},
78
  ],
79
+ temperature=0.0,
80
  )
81
  raw_response = (completion.choices[0].message.content or "").strip()
82
 
 
83
  action_dict = json.loads(raw_response)
84
  action = BlockchainAction(**action_dict)
85
  result = env.step(action)
86
 
87
  reward = result.reward
88
  error = None
89
+ action_str = json.dumps(action_dict).replace(" ", "")
90
 
91
  except Exception as e:
92
+ # FIX 2: Never output negative numbers or raw text! Bot parses this!
93
+ reward = 0.01
94
+ error = "format_error" # No spaces allowed in error logs
95
+ action_str = "{}" # Must be valid JSON brackets
96
+ result.done = True
97
 
98
+ # FIX 3: Strictly clamp reward before printing to the terminal
99
+ reward = max(0.01, min(0.99, reward))
100
  rewards.append(reward)
101
  steps_taken = step
102
 
 
105
  if result.done:
106
  break
107
 
108
+ score = sum(rewards) / steps_taken if steps_taken > 0 else 0.01
109
+ score = max(0.01, min(0.99, score))
 
110
  success = score >= 0.6
111
 
112
  finally: