app.py

import streamlit as st
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
import os
import os.path 
import pickle
import torch

base_model_id = "meta-llama/Llama-3.2-3B-Instruct"
model_id = "HiGenius/Headline-Generation-Model"
hf_token = os.environ.get('HF_TOKEN')

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

@st.cache_resource
def load_model():
    base_model = AutoModelForCausalLM.from_pretrained(base_model_id, use_auth_token=hf_token)
    model = PeftModel.from_pretrained(base_model, model_id).to(device)
    tokenizer = AutoTokenizer.from_pretrained(base_model_id, use_auth_token=hf_token)
    tokenizer.pad_token = tokenizer.eos_token
    tokenizer.padding_side='left'

    return tokenizer, model

tokenizer, model = load_model()

guideline_path = "./guidelines.txt"
with open(guideline_path, 'r', encoding='utf-8') as f:
    guidelines = f.read()

def process_prompt(tokenizer, content, video_summary = '', guidelines = None):
    if guidelines:
        system_prompt = "You are a helpful assistant that writes engaging headlines. To maximize engagement, you may follow these proven guidelines:\n" + guidelines
    else:
        system_prompt = "You are a helpful assistant that writes engaging headlines."

    user_prompt = (
        f"Below is an article and its accompanying video summary:\n\n"
        f"Article Content:\n{content}\n\n"
        f"Video Summary:\n{'None' if video_summary == '' else video_summary}\n\n"
        f"Write ONLY a single engaging headline that accurately reflects the article. Do not include any additional text, explanations, or options."
    )
    messages = [
        {"role": "system", "content": system_prompt},
        {"role": "user", "content": user_prompt},
    ]
    prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    return prompt


st.title("Article Headline Writer")
st.write("Write a catchy headline from content and video summary.")

# Inputs for content and video summary
content = st.text_area("Enter the article content:", placeholder="Type the main content of the article here...")
video_summary = st.text_area("Enter the summary of the article's accompanying video (optional):", placeholder="Type the summary of the video related to the article...")

if st.button("Generate Headline"):
    if content.strip():
        if not video_summary.strip():
            video_summary = ''
        prompt = process_prompt(tokenizer, content, video_summary, guidelines)
        inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=1024).to(device)
        
        st.write("### Generated 5 Potential Headlines:")
        for i in range(5):
            st.write(f"### Headline {i+1}")
            outputs = model.generate(**inputs,
                                   max_new_tokens=60,
                                   num_return_sequences=1, 
                                   do_sample=True,
                                   temperature=0.7)
            response = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True)
            response = response.replace('"', '')
            st.write(f"{response}")
    else:
        st.write("Please enter a valid prompt.")