v3
Browse files- inference.py +40 -42
inference.py
CHANGED
|
@@ -15,8 +15,6 @@ TASK_NAME = "customer-support"
|
|
| 15 |
BENCHMARK = "openenv"
|
| 16 |
|
| 17 |
MAX_STEPS = 20
|
| 18 |
-
MAX_TOTAL_REWARD = 20
|
| 19 |
-
SUCCESS_SCORE_THRESHOLD = 0.6
|
| 20 |
|
| 21 |
|
| 22 |
# LLM AGENT
|
|
@@ -38,7 +36,7 @@ refund / replace / reject / ask_proof
|
|
| 38 |
)
|
| 39 |
return response.choices[0].message.content.strip().lower()
|
| 40 |
except Exception as e:
|
| 41 |
-
print(f"[ERROR] LLM
|
| 42 |
return "reject"
|
| 43 |
|
| 44 |
|
|
@@ -64,12 +62,8 @@ def log_end(**kwargs):
|
|
| 64 |
print("[END]", kwargs, flush=True)
|
| 65 |
|
| 66 |
|
| 67 |
-
# MAIN
|
| 68 |
async def main():
|
| 69 |
-
rewards = []
|
| 70 |
-
steps_taken = 0
|
| 71 |
-
success = False
|
| 72 |
-
|
| 73 |
client = OpenAI(
|
| 74 |
base_url=API_BASE_URL,
|
| 75 |
api_key=API_KEY
|
|
@@ -77,59 +71,63 @@ async def main():
|
|
| 77 |
|
| 78 |
log_start(task=TASK_NAME, env=BENCHMARK, model=MODEL_NAME)
|
| 79 |
|
| 80 |
-
|
| 81 |
-
res = safe_post(f"{SPACE_URL}/reset")
|
| 82 |
-
result = res.json()
|
| 83 |
-
|
| 84 |
-
state = result["observation"]["echoed_message"]
|
| 85 |
-
done = result["done"]
|
| 86 |
-
|
| 87 |
-
for step in range(1, MAX_STEPS + 1):
|
| 88 |
-
if done:
|
| 89 |
-
break
|
| 90 |
-
|
| 91 |
-
message = get_model_message(client, state)
|
| 92 |
|
|
|
|
|
|
|
|
|
|
| 93 |
res = safe_post(
|
| 94 |
-
f"{SPACE_URL}/
|
| 95 |
-
json={"
|
| 96 |
)
|
| 97 |
result = res.json()
|
| 98 |
|
| 99 |
-
reward = result.get("reward", 0.0)
|
| 100 |
-
done = result.get("done", False)
|
| 101 |
state = result["observation"]["echoed_message"]
|
|
|
|
| 102 |
|
| 103 |
-
|
| 104 |
-
|
|
|
|
| 105 |
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 113 |
|
| 114 |
-
|
|
|
|
|
|
|
| 115 |
|
| 116 |
-
|
| 117 |
-
score = max(0.01, min(0.99, score))
|
| 118 |
|
| 119 |
-
|
| 120 |
|
| 121 |
except Exception as e:
|
| 122 |
print(f"[ERROR] {e}")
|
| 123 |
|
| 124 |
finally:
|
| 125 |
log_end(
|
| 126 |
-
success=
|
| 127 |
-
steps=
|
| 128 |
-
score=
|
| 129 |
-
rewards=
|
| 130 |
)
|
| 131 |
|
| 132 |
|
| 133 |
-
#
|
| 134 |
if __name__ == "__main__":
|
| 135 |
asyncio.run(main())
|
|
|
|
| 15 |
BENCHMARK = "openenv"
|
| 16 |
|
| 17 |
MAX_STEPS = 20
|
|
|
|
|
|
|
| 18 |
|
| 19 |
|
| 20 |
# LLM AGENT
|
|
|
|
| 36 |
)
|
| 37 |
return response.choices[0].message.content.strip().lower()
|
| 38 |
except Exception as e:
|
| 39 |
+
print(f"[ERROR] LLM failed: {e}")
|
| 40 |
return "reject"
|
| 41 |
|
| 42 |
|
|
|
|
| 62 |
print("[END]", kwargs, flush=True)
|
| 63 |
|
| 64 |
|
| 65 |
+
# MAIN
|
| 66 |
async def main():
|
|
|
|
|
|
|
|
|
|
|
|
|
| 67 |
client = OpenAI(
|
| 68 |
base_url=API_BASE_URL,
|
| 69 |
api_key=API_KEY
|
|
|
|
| 71 |
|
| 72 |
log_start(task=TASK_NAME, env=BENCHMARK, model=MODEL_NAME)
|
| 73 |
|
| 74 |
+
task_scores = []
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 75 |
|
| 76 |
+
try:
|
| 77 |
+
for task in ["easy", "medium", "hard"]:
|
| 78 |
+
# RESET TASK
|
| 79 |
res = safe_post(
|
| 80 |
+
f"{SPACE_URL}/reset",
|
| 81 |
+
json={"task": task}
|
| 82 |
)
|
| 83 |
result = res.json()
|
| 84 |
|
|
|
|
|
|
|
| 85 |
state = result["observation"]["echoed_message"]
|
| 86 |
+
done = result["done"]
|
| 87 |
|
| 88 |
+
for step in range(1, MAX_STEPS + 1):
|
| 89 |
+
if done:
|
| 90 |
+
break
|
| 91 |
|
| 92 |
+
message = get_model_message(client, state)
|
| 93 |
+
|
| 94 |
+
res = safe_post(
|
| 95 |
+
f"{SPACE_URL}/step",
|
| 96 |
+
json={"message": message}
|
| 97 |
+
)
|
| 98 |
+
result = res.json()
|
| 99 |
+
|
| 100 |
+
state = result["observation"]["echoed_message"]
|
| 101 |
+
done = result["done"]
|
| 102 |
+
|
| 103 |
+
log_step(
|
| 104 |
+
step=step,
|
| 105 |
+
action=message,
|
| 106 |
+
reward=result.get("reward", 0.0),
|
| 107 |
+
done=done,
|
| 108 |
+
error=None
|
| 109 |
+
)
|
| 110 |
|
| 111 |
+
# GET TASK SCORE FROM ENV
|
| 112 |
+
score = result.get("score", 0.5)
|
| 113 |
+
score = max(0.01, min(0.99, score))
|
| 114 |
|
| 115 |
+
task_scores.append(score)
|
|
|
|
| 116 |
|
| 117 |
+
final_score = sum(task_scores) / len(task_scores)
|
| 118 |
|
| 119 |
except Exception as e:
|
| 120 |
print(f"[ERROR] {e}")
|
| 121 |
|
| 122 |
finally:
|
| 123 |
log_end(
|
| 124 |
+
success=True,
|
| 125 |
+
steps=len(task_scores),
|
| 126 |
+
score=final_score if 'final_score' in locals() else 0.0,
|
| 127 |
+
rewards=task_scores
|
| 128 |
)
|
| 129 |
|
| 130 |
|
| 131 |
+
# ENTRY
|
| 132 |
if __name__ == "__main__":
|
| 133 |
asyncio.run(main())
|