app.py

import os
os.system("pip install gradio==4.44.1")
import gradio as gr
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline
from datasets import concatenate_datasets, load_dataset
import gc
from peft import PeftModel, PeftConfig
from langchain.chains import RetrievalQA
from langchain_community.vectorstores import FAISS
from langchain.embeddings import OpenAIEmbeddings
from langchain.docstore.document import Document
from langchain.llms import HuggingFacePipeline
from langchain.embeddings import HuggingFaceEmbeddings
from langchain.chains.question_answering import load_qa_chain
from langchain.prompts import PromptTemplate
import torch
import random
from langchain.document_loaders import WebBaseLoader
from langchain.text_splitter import CharacterTextSplitter
from langchain.memory import ConversationBufferMemory
import requests
import re

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Load Samsum dataset for generating questions
train_dataset = load_dataset("samsum", split='train', trust_remote_code=True)
val_dataset = load_dataset("samsum", split='validation', trust_remote_code=True)
samsum_dataset = concatenate_datasets([train_dataset, val_dataset])

model_name = "google/flan-t5-base"
tokenizer = AutoTokenizer.from_pretrained(model_name)
base_model = AutoModelForSeq2SeqLM.from_pretrained(model_name, torch_dtype=torch.bfloat16).to(device)
rlhf_model_path = "raghav-gaggar/PEFT_RLHF_TextSummarizer"
config = PeftConfig.from_pretrained(rlhf_model_path)
ppo_model = PeftModel.from_pretrained(base_model, rlhf_model_path).to(device)
merged_model = ppo_model.merge_and_unload().to(device)

base_model.eval()
ppo_model.eval()
merged_model.eval()

dialogsum_dataset = load_dataset("knkarthick/dialogsum", trust_remote_code=True)

def format_dialogsum_as_document(example):
    return Document(page_content=f"Dialogue:\n {example['dialogue']}\n\nSummary: {example['summary']}")

# Create documents from DialogSum dataset
documents = []
for split in ['train', 'validation', 'test']:
    documents.extend([format_dialogsum_as_document(example) for example in dialogsum_dataset[split]])

# Split the documents into chunks
text_splitter = CharacterTextSplitter(chunk_size=5200, chunk_overlap=0)
docs = text_splitter.split_documents(documents)

# Create embeddings and vector store for DialogSum documents
embeddings = HuggingFaceEmbeddings(
    model_name="sentence-transformers/all-MiniLM-L6-v2",
    model_kwargs={"device": "cuda" if torch.cuda.is_available() else "cpu"},
    encode_kwargs={"batch_size": 32}
)

vector_store = FAISS.from_documents(docs, embeddings)

# Initialize retriever for DialogSum documents
retriever = vector_store.as_retriever(search_kwargs={"k": 1})

prompt_template = """
Concisely summarize the dialogue in the end, like the example provided -

Example -
{context}

Dialogue to be summarized:
{question}

Summary:"""

PROMPT = PromptTemplate(
    template=prompt_template, input_variables=["context", "question"]
)

# Create a Hugging Face pipeline
summarization_pipeline = pipeline(
    "summarization",
    model=merged_model,
    tokenizer=tokenizer,
    max_length=150,
    min_length=20,
    do_sample=False,
)

# Wrap the pipeline in a LangChain LLM
llm = HuggingFacePipeline(pipeline=summarization_pipeline)

qa_chain = RetrievalQA.from_chain_type(
    llm, retriever=retriever, chain_type_kwargs={"prompt": PROMPT}
)

# Function for Gradio interface
def summarize_conversation(question):
    result = qa_chain({"query": question})
    return result["result"]

examples = [["Amanda: I baked cookies. Do you want some? \nJerry: Sure! \nAmanda: I'll bring you tomorrow :-)"]]

# Create Gradio interface
iface = gr.Interface(
    fn=summarize_conversation,
    inputs=gr.Textbox(lines=10, label="Enter conversation here"),
    outputs=gr.Textbox(label="Summary"),
    title="Conversation Summarizer",
    description="Enter a conversation, and the AI will provide a concise summary.",
    examples = examples
)

# Launch the app
iface.launch()