app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
import json
import os
from typing import Dict, Union

# --- Model and Instruction Configuration ---
MODEL_ID = "meta-llama/Llama-3.2-1B-Instruct"

SYSTEM_INSTRUCTION = """
You are a strict grading assistant.
Return ONLY a JSON object with:
- accuracy (float 0-10)
- grade (string A-D)
- feedback (string)
"""
# ------------------------------------------

# Load Model and Tokenizer once for the entire application
try:
    print(f"Loading model {MODEL_ID} for Gradio app...")
    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
    model = AutoModelForCausalLM.from_pretrained(
        MODEL_ID,
        torch_dtype=torch.bfloat16,
        device_map="auto",
    )
    TERMINATORS = [
        tokenizer.eos_token_id,
        tokenizer.convert_tokens_to_ids("<|eot_id|>") 
    ]
    MODEL_LOADED = True
except Exception as e:
    print(f"Error loading model or tokenizer: {e}")
    print("Gradio will run, but the grading function will return an error.")
    MODEL_LOADED = False
    tokenizer, model, TERMINATORS = None, None, None


def grade_response(student_response: str) -> Union[Dict, str]:
    """
    Core grading function (same as before)
    """
    if not MODEL_LOADED:
        return {"accuracy": 0.0, "grade": "Error", "feedback": "Model failed to load. Check console for details."}
    
    # 1. Construct the Message List
    messages = [
        {"role": "system", "content": SYSTEM_INSTRUCTION},
        {"role": "user", "content": f"Student response to grade: '{student_response}'"},
    ]

    # 2. Apply Chat Template and Tokenize
    input_ids = tokenizer.apply_chat_template(
        messages,
        add_generation_prompt=True,
        return_tensors="pt"
    ).to(model.device)

    # 3. Generate the Output
    try:
        output_ids = model.generate(
            input_ids,
            max_new_tokens=200,
            eos_token_id=TERMINATORS,
            do_sample=True,
            temperature=0.5,
            top_p=0.9,
        )
    except Exception as e:
        return {"accuracy": 0.0, "grade": "Error", "feedback": f"Generation error: {e}"}

    # 4. Decode the Raw Response
    raw_response = tokenizer.decode(
        output_ids[0][input_ids.shape[-1]:], 
        skip_special_tokens=True
    ).strip()
    
    # 5. Parse the JSON Output
    try:
        start_index = raw_response.find('{')
        end_index = raw_response.rfind('}') + 1
        json_string = raw_response[start_index:end_index]
        return json.loads(json_string)
        
    except json.JSONDecodeError:
        # If parsing fails, return a structured error response
        return {"accuracy": 0.0, "grade": "Error", "feedback": f"JSON Decode Error. Raw: {raw_response[:200]}..."}

# --- Gradio Wrapper Function ---

def gradio_grade_wrapper(student_response: str) -> tuple[float, str, str]:
    """
    Wraps the core grading function to match the required Gradio outputs.
    """
    result = grade_response(student_response)
    
    # Check if the result is a dictionary (the expected structured output)
    if isinstance(result, dict):
        # Gradio outputs: (accuracy, grade, feedback)
        return (
            result.get("accuracy", 0.0), 
            result.get("grade", "N/A"), 
            result.get("feedback", "No feedback generated.")
        )
    else:
        # Should not happen if error handling in grade_response is correct,
        # but here for extreme robustness.
        return (0.0, "ERROR", str(result))


# --- Gradio Interface Definition ---

with gr.Blocks(theme=gr.themes.Soft(), title="LLM Essay Grader") as demo:
    gr.Markdown("# 📝 LLM Essay Grading Assistant (Llama-3.2-1B-Instruct)")
    gr.Markdown(
        "Enter a student's response below to receive an automated grade, accuracy score, and feedback "
        "from the Llama-3.2-1B-Instruct model."
    )
    
    # Input Component
    with gr.Row():
        student_input = gr.Textbox(
            label="Student Response to Grade",
            placeholder="E.g., 'The main causes of the World War 2 were economic depression and poor leadership.'",
            lines=5,
            scale=3
        )
        grade_button = gr.Button("Submit for Grading", scale=1, variant="primary")

    gr.Markdown("---")
    gr.Markdown("## Grading Results")

    # Output Components arranged in a Row for visual clarity
    with gr.Row():
        accuracy_output = gr.Number(label="Accuracy (0-10)", interactive=False, precision=1)
        grade_output = gr.Textbox(label="Grade (A-D)", interactive=False)
        
    feedback_output = gr.Textbox(
        label="Detailed Feedback",
        interactive=False,
        lines=4,
        max_lines=10
    )

    # Event Listener: Connect the button click to the wrapper function
    grade_button.click(
        fn=gradio_grade_wrapper,
        inputs=[student_input],
        outputs=[accuracy_output, grade_output, feedback_output]
    )
    
    # Add Examples
    gr.Examples(
        examples=[
            ["The Earth is a cube and its main moon is Mars, which proves that gravity is fake."],
            ["A proper noun is a name used to designate a single, specific person, place, or thing, and is always capitalized."],
            ["The two main drivers of climate change are the burning of fossil fuels (releasing greenhouse gases) and deforestation."],
        ],
        inputs=student_input,
    )

# Launch the Gradio App
if __name__ == "__main__":
    demo.launch()