update

app.py

@@ -30,13 +30,16 @@ DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
 def build_agent():
     """Build a smolagents CodeAgent equipped for GAIA benchmark tasks."""
 
-    # Model — Use the HF Inference API (free tier)
-    # Fallback endpoint if the main model is overloaded:
-    # model_id='https://pflgm2locj2t89co.us-east-1.aws.endpoints.huggingface.cloud'
+    # Model — Use the HF Inference API
+    # Try multiple models in order of preference
+    model_id = os.getenv(
+        "MODEL_ID",
+        "Qwen/Qwen2.5-Coder-32B-Instruct"
+    )
     model = ModelClass(
         max_tokens=4096,
         temperature=0.1,
-        model_id="Qwen/Qwen2.5-Coder-32B-Instruct",
+        model_id=model_id,
         custom_role_conversions=None,
     )
 
@@ -61,29 +64,52 @@ def build_agent():
     return agent
 
 
-def extract_answer(raw_answer: str) -> str:
-    """Clean agent output to extract only the final answer value."""
+def extract_answer(raw_answer) -> str:
+    """Aggressively clean agent output to extract only the final answer value."""
     if raw_answer is None:
         return ""
     answer = str(raw_answer).strip()
 
-    # Remove common prefixes the agent might include
+    # If the answer contains final_answer("..."), extract the argument
+    fa_match = re.search(r'final_answer\(["\'](.+?)["\']\)', answer, re.DOTALL)
+    if fa_match:
+        answer = fa_match.group(1).strip()
+
+    # Remove code blocks (```py ... ```)
+    answer = re.sub(r'```[\s\S]*?```', '', answer).strip()
+
+    # Remove <end_code> tags and surrounding artifacts
+    answer = re.sub(r'<end_code>.*', '', answer, flags=re.DOTALL).strip()
+
+    # Remove Calling tools: [...] JSON metadata
+    answer = re.sub(r'Calling tools:.*', '', answer, flags=re.DOTALL).strip()
+
+    # Remove "Using the `final_answer` tool:" and similar
+    answer = re.sub(r'Using the `final_answer` tool:.*', '', answer, flags=re.DOTALL).strip()
+
+    # Remove Thought: / Code: sections if they leaked through
+    answer = re.sub(r'^Thought:.*?(?=\S)', '', answer, flags=re.DOTALL).strip()
+
+    # Remove common prefixes
     prefixes = [
-        "FINAL ANSWER:",
-        "Final Answer:",
-        "final answer:",
-        "The final answer is:",
-        "The answer is:",
-        "Answer:",
+        "FINAL ANSWER:", "Final Answer:", "final answer:",
+        "The final answer is:", "The final answer is ",
+        "The answer is:", "The answer is ",
+        "Answer:", "Final answer:",
     ]
     for prefix in prefixes:
-        if answer.startswith(prefix):
+        if answer.lower().startswith(prefix.lower()):
             answer = answer[len(prefix):].strip()
 
     # Remove surrounding quotes if present
-    if (answer.startswith('"') and answer.endswith('"')) or \
-       (answer.startswith("'") and answer.endswith("'")):
-        answer = answer[1:-1].strip()
+    if len(answer) >= 2:
+        if (answer[0] == '"' and answer[-1] == '"') or \
+           (answer[0] == "'" and answer[-1] == "'"):
+            answer = answer[1:-1].strip()
+
+    # Remove trailing periods (unless it's a decimal number)
+    if answer.endswith('.') and not re.match(r'^\d+\.$', answer):
+        answer = answer[:-1].strip()
 
     return answer
 
@@ -150,7 +176,7 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
             print(f"\n{'='*60}")
             print(f"Question {i+1}/{len(questions_data)} (task_id: {task_id})")
             print(f"Q: {question_text[:100]}...")
-            raw_answer = agent.run(question_text)
+            raw_answer = agent.run(question_text, reset=True)
             submitted_answer = extract_answer(raw_answer)
             print(f"A: {submitted_answer}")
             answers_payload.append({