app.py

import gradio as gr
from gradio.components import Textbox, Checkbox
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, T5ForConditionalGeneration
from peft import PeftModel
import torch
import datasets
from sentence_transformers import CrossEncoder
import math
import re
from nltk import sent_tokenize, word_tokenize
import nltk
nltk.download('punkt')

# Load cross encoder
top_k = 10
cross_encoder = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')

# Load your fine-tuned model and tokenizer
model_name = "google/flan-t5-large"
peft_name = "legacy107/flan-t5-large-ia3-covidqa"
tokenizer = AutoTokenizer.from_pretrained(model_name)
pretrained_model = T5ForConditionalGeneration.from_pretrained("google/flan-t5-large")
model = T5ForConditionalGeneration.from_pretrained("google/flan-t5-large")
model = PeftModel.from_pretrained(model, peft_name)

peft_name = "legacy107/flan-t5-large-ia3-bioasq-paraphrase"
paraphrase_model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
paraphrase_model = PeftModel.from_pretrained(paraphrase_model, peft_name)

max_length = 512
max_target_length = 200

# Load your dataset
dataset = datasets.load_dataset("minh21/COVID-QA-Chunk-64-testset-biencoder-data-90_10", split="train")
# dataset = dataset.shuffle()
dataset = dataset.select([4, 16, 22, 154, 63, 19, 31, 97, 133, 183])

# Context chunking
min_sentences_per_chunk = 3
chunk_size = 64
window_size = math.ceil(min_sentences_per_chunk * 0.25)
over_lap_chunk_size = chunk_size * 0.25

def chunk_splitter(context):
    sentences = sent_tokenize(context)
    chunks = []
    current_chunk = []

    for sentence in sentences:
        if len(current_chunk) < min_sentences_per_chunk:
            current_chunk.append(sentence)
            continue
        elif len(word_tokenize(' '.join(current_chunk) + " " + sentence)) < chunk_size:
            current_chunk.append(sentence)
            continue

        chunks.append(' '.join(current_chunk))
        new_chunk = current_chunk[-window_size:]
        new_window = window_size
        buffer_new_chunk = new_chunk

        while len(word_tokenize(' '.join(new_chunk))) <= over_lap_chunk_size:
            buffer_new_chunk = new_chunk
            new_window += 1
            new_chunk = current_chunk[-new_window:]
            if new_window >= len(current_chunk):
               break

        current_chunk = buffer_new_chunk
        current_chunk.append(sentence)


    if current_chunk:
        chunks.append(' '.join(current_chunk))

    return chunks


def clean_data(text):
    # Extract abstract content
    index = text.find("\nAbstract: ")
    if index != -1:
        cleaned_text = text[index + len("\nAbstract: "):]
    else:
        cleaned_text = text  # If "\nAbstract: " is not found, keep the original text

    # Remove both http and https links using a regular expression
    cleaned_text = re.sub(r'(http(s|)\/\/:( |)\S+)|(http(s|):\/\/( |)\S+)', '', cleaned_text)


    # Remove DOI patterns like "doi:10.1371/journal.pone.0007211.s003"
    cleaned_text = re.sub(r'doi:( |)\w+', '', cleaned_text)

    # Remove the "(0.11 MB DOC)" pattern
    cleaned_text = re.sub(r'\(0\.\d+ MB DOC\)', '', cleaned_text)

    cleaned_text = re.sub(r'www\.\w+(.org|)', '', cleaned_text)

    return cleaned_text


def paraphrase_answer(question, answer, use_pretrained=False):
    # Combine question and context
    input_text = f"question: {question}. Paraphrase the answer to make it more natural answer: {answer}"

    # Tokenize the input text
    input_ids = tokenizer(
        input_text,
        return_tensors="pt",
        padding="max_length",
        truncation=True,
        max_length=max_length,
    ).input_ids

    # Generate the answer
    with torch.no_grad():
        if use_pretrained:
            generated_ids = pretrained_model.generate(input_ids=input_ids, max_new_tokens=max_target_length)
        else:
            generated_ids = paraphrase_model.generate(input_ids=input_ids, max_new_tokens=max_target_length)

    # Decode and return the generated answer
    paraphrased_answer = tokenizer.decode(generated_ids[0], skip_special_tokens=True)

    return paraphrased_answer


def retrieve_context(question, contexts):
    # cross-encoder
    hits = [{"corpus_id": i} for i in range(len(contexts))]
    cross_inp = [[question, contexts[hit["corpus_id"]]] for hit in hits]
    cross_scores = cross_encoder.predict(cross_inp, show_progress_bar=False)

    for idx in range(len(cross_scores)):
        hits[idx]["cross-score"] = cross_scores[idx]

    hits = sorted(hits, key=lambda x: x["cross-score"], reverse=True)

    return " ".join(
        [contexts[hit["corpus_id"]] for hit in hits[0:top_k]]
    ).replace("\n", " ")


# Define your function to generate answers
def generate_answer(question, context, ground, do_pretrained, do_natural, do_pretrained_natural):
    contexts = chunk_splitter(clean_data(context))
    retrieved_context = retrieve_context(question, contexts)
    ground_in_context = (retrieved_context.find(ground) != -1)
    
    # Combine question and context
    input_text = f"question: {question} context: {retrieved_context}"

    # Tokenize the input text
    input_ids = tokenizer(
        input_text,
        return_tensors="pt",
        padding="max_length",
        truncation=True,
        max_length=max_length,
    ).input_ids

    # Generate the answer
    with torch.no_grad():
        generated_ids = model.generate(input_ids=input_ids, max_new_tokens=max_target_length)

    # Decode and return the generated answer
    generated_answer = tokenizer.decode(generated_ids[0], skip_special_tokens=True)

    # Paraphrase answer
    paraphrased_answer = ""
    if do_natural:
        paraphrased_answer = paraphrase_answer(question, generated_answer)

    # Get pretrained model's answer
    pretrained_answer = ""
    if do_pretrained:
        with torch.no_grad():
            pretrained_generated_ids = pretrained_model.generate(input_ids=input_ids, max_new_tokens=max_target_length)
            pretrained_answer = tokenizer.decode(pretrained_generated_ids[0], skip_special_tokens=True)

    # Get pretrained model's natural answer
    pretrained_paraphrased_answer = ""
    if do_pretrained_natural:
        pretrained_paraphrased_answer = paraphrase_answer(question, generated_answer, True)

    return generated_answer, paraphrased_answer, ("Yes" if ground_in_context else "No"), pretrained_answer, pretrained_paraphrased_answer, retrieved_context


# Define a function to list examples from the dataset
def list_examples():
    examples = []
    for example in dataset:
        context = example["context"]
        question = example["question"]
        answer = example["answer"]
        examples.append([question, context, answer, True, True, True])
    return examples


# Create a Gradio interface
iface = gr.Interface(
    fn=generate_answer,
    inputs=[
        Textbox(label="Question"),
        Textbox(label="Context"),
        Textbox(label="Ground truth"),
        Checkbox(label="Include pretrained model's result"),
        Checkbox(label="Include natural answer"),
        Checkbox(label="Include pretrained model's natural answer")
    ],
    outputs=[
        Textbox(label="Generated Answer"),
        Textbox(label="Natural Answer"),
        Textbox(label="Ground truth in the retrieved context"),
        Textbox(label="Pretrained Model's Answer"),
        Textbox(label="Pretrained Model's Natural Answer"),
        Textbox(label="Retrieved Context")
    ],
    examples=list_examples(),
    examples_per_page=1,
)

# Launch the Gradio interface
iface.launch()