app.py

from unsloth import FastLanguageModel
import torch
import gradio as gr
import xml.etree.ElementTree as ET
import re

"""
This module provides utilities for extracting structured data from text blocks.
It supports parsing XML-like structures, Markdown-like formatting, and alternative
text representations for extracting "choice" and "justification" fields.
Functions:
    extract_from_xml_et(text: str) -> dict
        Parses XML-like text and extracts key-value pairs.
    extract_choice(text: str) -> str
        Extracts the choice (e.g., A), B), C), D)) from a text block.
    extract_justification(text: str) -> str
        Extracts the justification text from a text block.
    extract_from_markdown_regex(text: str) -> dict
        Extracts data from Markdown-like structured text, specifically "choice"
        and "justification" fields.
    extract_fields(text: str) -> list
        Processes text blocks to extract structured data using a combination of
        XML parsing, regex-based choice and justification extraction, and Markdown-like parsing.
"""

def extract_from_xml_et(text: str) -> dict:
    """
    Parses an XML-like string and extracts key-value pairs from its elements.
    Parameters:
        text (str): A string containing XML-like content (e.g., <tag>value</tag>).
    Returns:
        dict: A dictionary where the keys are lowercase XML tags and the values
              are their corresponding text content.
        None: Returns None if XML parsing fails.
    Example:
        >>> text = '<key>"value"</key>'
        >>> extract_from_xml_et(text)
        {'key': 'value'}
    """
    try:
        wrapped_text = f"<root>{text}</root>"
        root = ET.fromstring(wrapped_text)
        data = {}
        for child in root:
            if child.text:
                value = child.text.strip().strip('"')
                data[child.tag.lower()] = value
        return data
    except ET.ParseError:
        return None


def extract_choice(text: str) -> str:
    """
    Extracts the choice (e.g., A), B), C), D)) from a text block.
    Parameters:
        text (str): Input text to search for the choice.
    Returns:
        str: The extracted choice, or None if not found.
    Example:
        >>> text = "A) This is a sample choice."
        >>> extract_choice(text)
        'A)'
    """
    choice_pattern = r'([A-D]\))'
    match = re.search(choice_pattern, text)
    if match:
        return match.group(1).strip()
    return None


def extract_justification(text: str) -> str:
    """
    Extracts the justification text from a text block.
    Parameters:
        text (str): Input text to search for the justification.
    Returns:
        str: The extracted justification, or None if not found.
    Example:
        >>> text = "- Justification: This is the reason."
        >>> extract_justification(text)
        'This is the reason.'
    """
    justification_pattern = r'(?:- )?Justification:\s*(.+)'
    match = re.search(justification_pattern, text)
    if match:
        return match.group(1).strip()
    return None


def extract_from_markdown_regex(text: str) -> dict:
    """
    Extracts structured data from Markdown-like text blocks.
    Parameters:
        text (str): Input text containing Markdown-like content, with **choice**
                    and **justification** fields.
    Returns:
        dict: A dictionary containing "choice" and "justification", or None if no match is found.
    Example:
        >>> text = "**choice**: A **justification**: This is the reason."
        >>> extract_from_markdown_regex(text)
        {'choice': 'A', 'justification': 'This is the reason.'}
    """
    choice_pattern = r'\*\*choice\*\*:\s*(.+?)'
    justification_pattern = r'\*\*justification\*\*:\s*([\s\S]+?)(?=\*\*choice\*\*|$)'
    choice_match = re.search(choice_pattern, text)
    justification_match = re.search(justification_pattern, text)

    if choice_match and justification_match:
        return {
            "choice": choice_match.group(1).strip(),
            "justification": justification_match.group(1).strip()
        }
    return None


def extract_fields(text: str) -> list:
    """
    Processes text blocks to extract structured data.
    This function attempts to parse each block using the following methods:
        1. XML Parsing: Uses extract_from_xml_et to handle XML-like content.
        2. Regex for Choice and Justification: Extracts these fields separately.
        3. Markdown Parsing: Uses extract_from_markdown_regex for Markdown-like structures.
    Parameters:
        text (str): Input text containing one or more blocks of data.
    Returns:
        list: A list of dictionaries, each containing extracted data from a block.
    Workflow:
        1. Splits the input text into blocks using double line breaks (\n\n).
        2. For each block:
            - Attempts to parse it using extract_from_xml_et.
            - If unsuccessful, tries extract_choice and extract_justification.
            - Finally, falls back to extract_from_markdown_regex.
        3. Aggregates the results into a list of dictionaries.
    Example:
        >>> text = '''
        <key>"value"</key>
        **choice**: A **justification**: This is the reason.
        A) Taking all reasonable measures to safeguard user data,
            - Justification: This is the reason.
        '''
        >>> extract_fields(text)
        [
            {'key': 'value'},
            {'choice': 'A', 'justification': 'This is the reason.'},
            {'choice': 'A)', 'justification': 'This is the reason.'}
        ]
    """
    entries = []
    blocks = re.split(r'\n\s*\n', text.strip())  # Split text into blocks by double newlines

    for block in blocks:
        print("Processing Block:", block)
        extracted_data = {}

        # Try extracting using XML
        xml_data = extract_from_xml_et(block)
        if xml_data:
            print("Extracted via XML:", xml_data)
            entries.append(xml_data)
            continue

        # Try extracting using separated choice and justification regex
        choice = extract_choice(block)
        justification = extract_justification(block)
        if choice or justification:
            extracted_data["choice"] = choice
            extracted_data["justification"] = justification
            entries.append(extracted_data)
            continue

        # Try extracting using Markdown regex
        markdown_data = extract_from_markdown_regex(block)
        if markdown_data:
            print("Extracted via Markdown Regex:", markdown_data)
            entries.append(markdown_data)

    return entries


### The code initializes the LLM model and tokenizer from a fine-tuned checkpoint located in a directory called unified_model.
model,tokenizer = FastLanguageModel.from_pretrained('./unified_model')


def generate_response_true_false(instruction):
    """
    Generates a response using your fine-tuned model based on the provided instruction.
    This function enables faster inference through the `FastLanguageModel` and prepares a
    prompt for the model to determine whether the given statement is "True" or "False".
    Args:
        instruction (str): A string containing the statement and instructions to be evaluated.
    Returns:
        str: "True" or "False" based on the model's response, or "Unable to determine" if the
             response cannot be parsed reliably.
    """
    FastLanguageModel.for_inference(model) # Enable native 2x faster inference within the function
    prompt = f"""### Instruction:
Determine if the following statement is true or false. Respond only with "True" or "False".
### Statement:
{instruction}
### Answer:"""

    inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
    with torch.no_grad():
        outputs = model.generate(**inputs, max_new_tokens=50)

    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    response = response.split("### Answer:")[-1].strip()

    # Extract True/False from response
    if response.lower() == "true":
        return "True"
    elif response.lower() == "false":
        return "False"
    else:
        # Try to identify the answer even if it's not perfectly formatted
        if "true" in response.lower():
            return "True"
        elif "false" in response.lower():
            return "False"
        else:
            return "Unable to determine."

def generate_response_open_ended(instruction):
    """
    Generates a response using your fine-tuned model based on the provided instruction.
    This function enables faster inference through the `FastLanguageModel` and prepares a
    prompt for the model to determine whether the given statement is "True" or "False".
    Args:
        instruction (str): A string containing the statement and instructions to be evaluated.
    Returns:
        str: A response from the model to the provided question or "Unable to determine" if the
             response cannot be parsed reliably.
    """
    FastLanguageModel.for_inference(model) # Enable native 2x faster inference within the function
    prompt = f"""### Instruction:
Answer the provided question with the knowledge provided to you
### Question:
{instruction}
### Answer:
"""

    inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
    with torch.no_grad():
        outputs = model.generate(**inputs,early_stopping=False,min_length=50,length_penalty=2,max_length=200)
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    # Extract the answer from the generated response by splitting on "### Answer:"
    response = response.split('### Answer:')[1]
    return response

def generate_response_multiple_choice(question,choice_A,choice_B,choice_C,choice_D):

    instruction = f'''{question}
  Choices:
  A) {choice_A},
  B) {choice_B},
  C) {choice_C},
  D) {choice_D}
    '''

    """
    Generates a response using a fine-tuned language model for multiple-choice questions.
    Args:
        instruction (str): A string containing the question and its options.
    Returns:
        dict: A dictionary with the selected choice and its justification.
              Example:
              {
                  "choice": "A",
                  "justification": "Explanation for why Option A is correct."
              }
              If the model fails to provide a valid response, defaults to:
              {
                  "choice": "None",
                  "justification": "Could not parse JSON"
              }
    """
    # Enable native faster inference for the model
    FastLanguageModel.for_inference(model)

    # Define the prompt with a detailed instruction for the model
    prompt = f"""### Instruction:
    In the following question, you are provided with 4 choices. Select the best choice based on the knowledge provided and provide a justification for that choice.
    **You must return only your response with the following keys:**
      - "choice": The best choice letter
      - "justification": The justification for your choice
    **Example Response:**
      **choice**: A
      **justification**: Explanation for why Option A is correct
    ### Question:
    {instruction}
    ### Answer:
    
    """

    # Tokenize the prompt and move it to GPU for inference
    inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

    # Generate a response from the model
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            early_stopping=True,
            min_length=50,
            length_penalty=2,
            do_sample=True,
            max_new_tokens=300,
            top_p=0.95,
            top_k=50,
            temperature=0.65,
            num_return_sequences=1
        )

    # Decode the response into text
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    # Extract the answer from the generated response by splitting on "### Answer:"
    response = response.split('### Answer:')[1]
    print("RESPONSE",response)
    data = extract_fields(response)
    if len(data) == 0:
      response = {"choice": data[0]['choice'], "justification": data[0]['justification']}
    else:
      response = {"choice": data[-1]['choice'], "justification": data[-1]['justification']}
    return response

def true_false_greet(question):
    if question == "":
        # Return a default response if no input is given
        return "No question was given to answer"
    else:
        # Call a placeholder function (must be implemented separately)
        response = generate_response_true_false(question)  # Note: This function is not defined in this code
        return f"{response}!"

def open_ended_greet(question):
    """
    Processes the user's question and returns a response.
    Args:
        question (str): The input text provided by the user.
    Returns:
        str: A processed response. If no input is given, a default message is returned.
    """
    if question == "":
        # Return a default response if no question is provided
        return "No question was given to answer"
    else:
        # Call a placeholder function (must be implemented separately) to generate a response
        response = generate_response_open_ended(question)  # Note: generate_response is not defined in this snippet

        # Return the formatted response
        return f"{response}!"

def multiple_choice_greet(question, choice_A, choice_B, choice_C, choice_D):
    """
    Processes the user's question and multiple-choice options to generate a response.
    Args:
        question (str): The input question provided by the user.
        choice_A (str): Option A for the question.
        choice_B (str): Option B for the question.
        choice_C (str): Option C for the question.
        choice_D (str): Option D for the question.
    Returns:
        str: A response based on the input.
             If no question is provided, returns a default message.
             If no choices are provided, returns a default message.
    """
    if question == "":
        # Return a default response if no question is provided
        return "No question was given to answer"
    if choice_A == "" and choice_B == "" and choice_C == "" and choice_D == "":
        # Return a default response if no choices are provided
        return "No choice was given"
    else:
        # Call a placeholder function (must be implemented separately) to generate a response
        response = generate_response_multiple_choice(question, choice_A, choice_B, choice_C, choice_D)
        actual_response = "Selected Choice: " + response['choice'] + "\nJustification: " + response['justification']
        # Return the formatted response
        return f"{actual_response}"

#### Function which enables the visibility of true/false questions interface
def show_true_false_interface():
    return gr.update(visible=True), gr.update(visible=False), gr.update(visible=False)

#### Function which enables the visibility of open-ended questions interface
def show_open_ended_interface():
    return gr.update(visible=False), gr.update(visible=True), gr.update(visible=False)

#### Function which enables the visibility of multiple-choice questions interface
def show_multiple_choice_interface():
    return gr.update(visible=False), gr.update(visible=False), gr.update(visible=True)

# print(generate_response_multiple_choice("Which of the following best describes a bank’s legal duty in cases of phishing, according to Greek law?",
#                                         "Taking all reasonable measures to safeguard user data and transactions",
#                                         "Ensuring absolute prevention of all cyberattacks",
#                                         "Holding customers solely responsible for phishing losses",
#                                         "Avoiding liability by implementing disclaimers"
#                                         ))


with gr.Blocks() as demo:

    ### We define a row in which we create the navigation buttons for each question type
    with gr.Row():
        btn_t_f = gr.Button('True/False questions')
        btn_open_ended = gr.Button('Open-Ended questions')
        btn_m_c = gr.Button('Multiple-Choice questions')

    ### We define the interface for the true/false questions
    with gr.Column(visible=True) as true_false_interface:
        gr.Markdown("## True-False Template")
        question_simple = gr.Textbox(label="Enter your question")
        simple_output = gr.Textbox(label="Output", interactive=False)
        submit_simple = gr.Button("Submit")
        submit_simple.click(true_false_greet, inputs=question_simple, outputs=simple_output)

    ### We define the interface for the open-ended questions
    with gr.Column(visible=False) as open_ended_interface:
        gr.Markdown("## Open Ended Template")
        question_simple = gr.Textbox(label="Enter your question")
        simple_output = gr.Textbox(label="Output", interactive=False)
        submit_simple = gr.Button("Submit")
        submit_simple.click(open_ended_greet, inputs=question_simple, outputs=simple_output)

    ### We define the interface for the multiple-choice questions
    with gr.Column(visible=False) as mc_interface:
        gr.Markdown("## Multiple-Choice Template")
        question_mc = gr.Textbox(label="Enter your question")
        choice_A = gr.Textbox(label="Choice A")
        choice_B = gr.Textbox(label="Choice B")
        choice_C = gr.Textbox(label="Choice C")
        choice_D = gr.Textbox(label="Choice D")
        mc_output = gr.Textbox(label="Output", interactive=False)
        submit_mc = gr.Button("Submit")
        submit_mc.click(multiple_choice_greet, inputs=[question_mc, choice_A, choice_B, choice_C, choice_D], outputs=mc_output)

    ### If a navigation button is clicked, a visibility function is executed
    btn_t_f.click(show_true_false_interface, outputs=[true_false_interface, open_ended_interface, mc_interface])
    btn_open_ended.click(show_open_ended_interface, outputs=[true_false_interface, open_ended_interface, mc_interface])
    btn_m_c.click(show_multiple_choice_interface, outputs=[true_false_interface, open_ended_interface, mc_interface])

demo.launch()