import os import requests import gradio as gr import spaces from transformers import pipeline from smolagents import CodeAgent, DuckDuckGoSearchTool, TransformersModel, tool from smolagents import InferenceClientModel, CodeAgent from huggingface_hub import InferenceClient # 1. Custom tool to download files if needed for tasks @tool def get_file_content_tool(task_id: str) -> str: """ Downloads files associated with a specific task ID from the course API. Args: task_id: The ID of the task requiring a file. """ url = f"https://agents-course-unit4-scoring.hf.space/files/{task_id}" response = requests.get(url) if response.status_code == 200: filename = f"temp_{task_id}" with open(filename, "wb") as f: f.write(response.content) return f"File downloaded successfully as {filename}." return f"Failed to download file for task {task_id}" # 2. Initialize model and agent # 2. Initialize model and agent # 2. Initialize model and agent with a free serverless-compatible model # 2. Initialize model and agent with a serverless-supported model model = InferenceClientModel( model_id="meta-llama/Llama-3.3-70B-Instruct", token=os.getenv("HF_TOKEN") ) agent = CodeAgent( tools=[DuckDuckGoSearchTool(), get_file_content_tool], model=model, max_steps=6, verbosity_level=1 ) # 3. Core evaluation function wrapped with @spaces.GPU @spaces.GPU def run_evaluation(username): if not username: return "Please log in with your Hugging Face account first." # Updated scoring API base URL api_base = "https://agents-course-unit4-scoring.hf.space" headers = {"User-Agent": "Mozilla/5.0"} try: resp = requests.get(f"{api_base}/questions", headers=headers, timeout=15) except Exception as e: return f"Connection error when reaching evaluation server: {str(e)}" if resp.status_code != 200: return f"Failed to fetch questions. Status code: {resp.status_code}. Response: {resp.text}" try: questions_data = resp.json() except Exception: return f"Invalid JSON response from server: {resp.text}" answers = [] log_output = "" for item in questions_data: task_id = item.get("task_id") question_text = item.get("question") log_output += f"\n--- Processing Task ID: {task_id} ---\nQuestion: {question_text}\n" try: prompt = f"{question_text}\nProvide ONLY the direct, final answer string with no conversational padding." raw_answer = agent.run(prompt) clean_answer = str(raw_answer).strip() answers.append({ "task_id": task_id, "submitted_answer": clean_answer }) log_output += f"Answer: {clean_answer}\n" except Exception as e: log_output += f"Error executing task {task_id}: {str(e)}\n" answers.append({ "task_id": task_id, "submitted_answer": "Error" }) # 4. Submit results to the course API code_link = f"https://huggingface.co/spaces/{username}/Final_Assignment_Template/tree/main" payload = { "username": username, "agent_code": code_link, "answers": answers } try: sub_resp = requests.post(f"{api_base}/submit", json=payload, headers=headers, timeout=20) if sub_resp.status_code == 200: result_json = sub_resp.json() log_output += f"\nSubmission Successful! Response: {result_json}" else: log_output += f"\nSubmission Failed (Status {sub_resp.status_code}): {sub_resp.text}" except Exception as e: log_output += f"\nSubmission Exception: {str(e)}" return log_output # 5. Build Gradio UI with gr.Blocks() as demo: gr.Markdown("# Custom Agent Evaluation Runner") gr.Markdown("Click below to run your custom `smolagents` code against the Unit 4 benchmark questions and submit automatically.") with gr.Row(): login_button = gr.LoginButton() with gr.Row(): submit_btn = gr.Button("Run Evaluation & Submit All Answers") output_box = gr.Textbox(label="Run Status / Submission Result", lines=15) submit_btn.click( fn=run_evaluation, inputs=[login_button], outputs=[output_box] ) if __name__ == "__main__": demo.launch()