Create gaia_submit.py
Browse files- gaia_submit.py +57 -0
gaia_submit.py
ADDED
@@ -0,0 +1,57 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import requests
|
2 |
+
import yaml
|
3 |
+
from smolagents import CodeAgent, HfApiModel
|
4 |
+
from tools.final_answer import FinalAnswerTool
|
5 |
+
|
6 |
+
GAIA_API = "https://gaia-api.spaces.huggingface.tech"
|
7 |
+
HF_USERNAME = "RoomSamurai"
|
8 |
+
SPACE_CODE_URL = f"https://huggingface.co/spaces/{HF_USERNAME}/AiAgentCourse_final_agent/tree/main"
|
9 |
+
|
10 |
+
with open("prompts.yaml", "r") as f:
|
11 |
+
prompt_templates = yaml.safe_load(f)
|
12 |
+
|
13 |
+
model = HfApiModel(model_id="Qwen/Qwen2.5-Coder-32B-Instruct", temperature=0.5, max_tokens=2096)
|
14 |
+
final_answer = FinalAnswerTool()
|
15 |
+
|
16 |
+
agent = CodeAgent(
|
17 |
+
model=model,
|
18 |
+
tools=[final_answer],
|
19 |
+
prompt_templates=prompt_templates,
|
20 |
+
max_steps=6
|
21 |
+
)
|
22 |
+
|
23 |
+
def get_questions():
|
24 |
+
r = requests.get(f"{GAIA_API}/questions")
|
25 |
+
return r.json()
|
26 |
+
|
27 |
+
def submit_all(answers):
|
28 |
+
payload = {
|
29 |
+
"username": HF_USERNAME,
|
30 |
+
"agent_code": SPACE_CODE_URL,
|
31 |
+
"answers": answers
|
32 |
+
}
|
33 |
+
r = requests.post(f"{GAIA_API}/submit", json=payload)
|
34 |
+
return r.json()
|
35 |
+
|
36 |
+
if __name__ == "__main__":
|
37 |
+
print("Fetching GAIA questions...")
|
38 |
+
questions = get_questions()
|
39 |
+
|
40 |
+
submitted = []
|
41 |
+
|
42 |
+
for q in questions[:20]:
|
43 |
+
print("\nQuestion:", q["question"])
|
44 |
+
try:
|
45 |
+
response = agent.run(q["question"])
|
46 |
+
except Exception as e:
|
47 |
+
print("Agent error:", e)
|
48 |
+
response = "error"
|
49 |
+
|
50 |
+
submitted.append({
|
51 |
+
"task_id": q["id"],
|
52 |
+
"submitted_answer": str(response).strip()
|
53 |
+
})
|
54 |
+
|
55 |
+
print("\nSubmitting answers...")
|
56 |
+
result = submit_all(submitted)
|
57 |
+
print("Submission result:", result)
|