app.py

import os
import gradio as gr
import logging
from transformers import MT5Tokenizer, MT5ForConditionalGeneration

# Setup logging
logging.basicConfig(level=logging.DEBUG, format='%(asctime)s - %(levelname)s - %(message)s')

# Load your fine-tuned mT5 model
model_name = "Addaci/mT5-small-experiment-13-checkpoint-2790"
tokenizer = MT5Tokenizer.from_pretrained(model_name)
model = MT5ForConditionalGeneration.from_pretrained(model_name)

def correct_htr(raw_htr_text):
    try:
        logging.info("Processing HTR correction...")
        inputs = tokenizer(raw_htr_text, return_tensors="pt", max_length=512, truncation=True)
        logging.debug(f"Tokenized Inputs for HTR Correction: {inputs}")

        outputs = model.generate(**inputs, max_length=128, num_beams=4, early_stopping=True)
        logging.debug(f"Generated Output (Tokens) for HTR Correction: {outputs}")

        corrected_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
        logging.debug(f"Decoded Output for HTR Correction: {corrected_text}")
        return corrected_text
    except Exception as e:
        logging.error(f"Error in HTR Correction: {e}", exc_info=True)
        return str(e)

def summarize_text(legal_text):
    try:
        logging.info("Processing summarization...")
        inputs = tokenizer("summarize: " + legal_text, return_tensors="pt", max_length=512, truncation=True)
        logging.debug(f"Tokenized Inputs for Summarization: {inputs}")

        outputs = model.generate(**inputs, max_length=150, num_beams=4, early_stopping=True)
        logging.debug(f"Generated Summary (Tokens): {outputs}")

        summary = tokenizer.decode(outputs[0], skip_special_tokens=True)
        logging.debug(f"Decoded Summary: {summary}")
        return summary
    except Exception as e:
        logging.error(f"Error in Summarization: {e}", exc_info=True)
        return str(e)

def answer_question(legal_text, question):
    try:
        logging.info("Processing question-answering...")
        formatted_input = f"question: {question} context: {legal_text}"
        inputs = tokenizer(formatted_input, return_tensors="pt", max_length=512, truncation=True)
        logging.debug(f"Tokenized Inputs for Question Answering: {inputs}")

        outputs = model.generate(**inputs, max_length=150, num_beams=4, early_stopping=True)
        logging.debug(f"Generated Answer (Tokens): {outputs}")

        answer = tokenizer.decode(outputs[0], skip_special_tokens=True)
        logging.debug(f"Decoded Answer: {answer}")
        return answer
    except Exception as e:
        logging.error(f"Error in Question Answering: {e}", exc_info=True)
        return str(e)

# Create the Gradio Blocks interface
with gr.Blocks() as demo:
    gr.Markdown("# mT5 Legal Assistant")
    gr.Markdown("Use this tool to correct raw HTR, summarize legal texts, or answer questions about legal cases.")

    # Adding external link buttons with a box around each and bold text
    with gr.Row():
        gr.HTML('''
            <div style="display: flex; gap: 10px;">
                <div style="border: 2px solid black; padding: 10px; display: inline-block;">
                    <a href="http://www.marinelives.org/wiki/Tools:_Admiralty_court_legal_glossary" target="_blank">
                        <button style="font-weight:bold;">Admiralty Court Legal Glossary</button>
                    </a>
                </div>
                <div style="border: 2px solid black; padding: 10px; display: inline-block;">
                    <a href="https://raw.githubusercontent.com/Addaci/HCA/refs/heads/main/HCA_13_70_Full_Volume_Processed_Text_EDITED_Ver.1.2_18062024.txt" target="_blank">
                        <button style="font-weight:bold;">HCA 13/70 Ground Truth (1654-55)</button>
                    </a>
                </div>
            </div>
        ''')

    with gr.Tab("Correct HTR"):
        gr.Markdown("### Correct Raw HTR Text")
        raw_htr_input = gr.Textbox(lines=5, placeholder="Enter raw HTR text here...")
        corrected_output = gr.Textbox(lines=5, placeholder="Corrected HTR text")
        correct_button = gr.Button("Correct HTR")
        clear_button = gr.Button("Clear")
        
        correct_button.click(correct_htr, inputs=raw_htr_input, outputs=corrected_output)
        clear_button.click(lambda: ("", ""), outputs=[raw_htr_input, corrected_output])

    with gr.Tab("Summarize Legal Text"):
        gr.Markdown("### Summarize Legal Text")
        legal_text_input = gr.Textbox(lines=10, placeholder="Enter legal text to summarize...")
        summary_output = gr.Textbox(lines=5, placeholder="Summary of legal text")
        summarize_button = gr.Button("Summarize Text")
        clear_button = gr.Button("Clear")
        
        summarize_button.click(summarize_text, inputs=legal_text_input, outputs=summary_output)
        clear_button.click(lambda: ("", ""), outputs=[legal_text_input, summary_output])

    with gr.Tab("Answer Legal Question"):
        gr.Markdown("### Answer a Question Based on Legal Text")
        legal_text_input_q = gr.Textbox(lines=10, placeholder="Enter legal text...")
        question_input = gr.Textbox(lines=2, placeholder="Enter your question...")
        answer_output = gr.Textbox(lines=5, placeholder="Answer to your question")
        answer_button = gr.Button("Get Answer")
        clear_button = gr.Button("Clear")
        
        answer_button.click(answer_question, inputs=[legal_text_input_q, question_input], outputs=answer_output)
        clear_button.click(lambda: ("", "", ""), outputs=[legal_text_input_q, question_input, answer_output])

# Launch the Gradio interface
demo.launch()