Upload app.py
Browse files
app.py
ADDED
@@ -0,0 +1,128 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# -*- coding: utf-8 -*-
|
2 |
+
"""Fake News Generator & Detector.ipynb
|
3 |
+
|
4 |
+
Automatically generated by Colab.
|
5 |
+
|
6 |
+
Original file is located at
|
7 |
+
https://colab.research.google.com/drive/1k_GVU6WO9Mggr87RxntJFq2S1M6R4dI0
|
8 |
+
|
9 |
+
# π° **Fake News Generator & Detector**
|
10 |
+
|
11 |
+
This educational and awareness-focused Generative AI application explores the dual power of AI in creating and detecting fake news. It allows users to generate fake news headlines based on a selected category (e.g., politics, tech), a subject or entity (e.g., Apple, Government), and number of headlines. It also provides a detection system where users can input a news snippet to check whether it is likely to be fake or real using a fine-tuned BERT model.
|
12 |
+
|
13 |
+
**The application is built using:**
|
14 |
+
|
15 |
+
Transformers library
|
16 |
+
(For GPT-2 based fake news generation and BERT-based fake news classification)
|
17 |
+
|
18 |
+
Gradio
|
19 |
+
(To create a clean, tabbed web interface for both generation and detection)
|
20 |
+
|
21 |
+
Google Colab / Python
|
22 |
+
(For backend development, model training/fine-tuning, and deployment)
|
23 |
+
|
24 |
+
**This project aims to:**
|
25 |
+
|
26 |
+
β¨ Demonstrate the capabilities and risks of large language models
|
27 |
+
|
28 |
+
π Raise awareness about misinformation in media
|
29 |
+
|
30 |
+
π§ Promote responsible AI development through practical exploration
|
31 |
+
|
32 |
+
This project showcases how AI can both contribute to and combat misinformation, serving as a tool for learning, experimentation, and ethics in AI."""
|
33 |
+
|
34 |
+
# app.py
|
35 |
+
|
36 |
+
import torch
|
37 |
+
import torch.nn.functional as F
|
38 |
+
from transformers import (
|
39 |
+
GPT2LMHeadModel,
|
40 |
+
GPT2Tokenizer,
|
41 |
+
AutoTokenizer,
|
42 |
+
AutoModelForSequenceClassification
|
43 |
+
)
|
44 |
+
import gradio as gr
|
45 |
+
|
46 |
+
# Device setup
|
47 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
48 |
+
|
49 |
+
# Load GPT-2 for Fake News Generation
|
50 |
+
gpt2_tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
|
51 |
+
gpt2_model = GPT2LMHeadModel.from_pretrained("gpt2").to(device)
|
52 |
+
gpt2_model.eval()
|
53 |
+
|
54 |
+
# Headline generation function
|
55 |
+
def generate_fake_headlines(category, subject, num_headlines=1, max_length=20):
|
56 |
+
headlines = []
|
57 |
+
for i in range(num_headlines):
|
58 |
+
prompt = f"Write a fake {category} news headline about {subject}:\n"
|
59 |
+
input_ids = gpt2_tokenizer.encode(prompt, return_tensors='pt').to(device)
|
60 |
+
|
61 |
+
output = gpt2_model.generate(
|
62 |
+
input_ids,
|
63 |
+
max_length=len(input_ids[0]) + max_length,
|
64 |
+
temperature=0.9,
|
65 |
+
top_p=0.9,
|
66 |
+
do_sample=True,
|
67 |
+
pad_token_id=gpt2_tokenizer.eos_token_id,
|
68 |
+
no_repeat_ngram_size=2,
|
69 |
+
early_stopping=True
|
70 |
+
)
|
71 |
+
|
72 |
+
generated_text = gpt2_tokenizer.decode(output[0], skip_special_tokens=True)
|
73 |
+
headline = generated_text.replace(prompt, "").strip().split("\n")[0]
|
74 |
+
headlines.append(headline)
|
75 |
+
return headlines
|
76 |
+
|
77 |
+
# Gradio function wrapper
|
78 |
+
def gradio_fake_news_generator(category, subject, num_headlines):
|
79 |
+
try:
|
80 |
+
num = int(num_headlines)
|
81 |
+
if num <= 0 or num > 10:
|
82 |
+
return "Please enter a number between 1 and 10."
|
83 |
+
except:
|
84 |
+
return "Invalid input for number of headlines."
|
85 |
+
|
86 |
+
headlines = generate_fake_headlines(category, subject, num)
|
87 |
+
return "\n".join([f"{i+1}. {h}" for i, h in enumerate(headlines)])
|
88 |
+
|
89 |
+
# Load BERT model for fake news detection (no token required)
|
90 |
+
bert_tokenizer = AutoTokenizer.from_pretrained("Pulk17/Fake-News-Detection")
|
91 |
+
bert_model = AutoModelForSequenceClassification.from_pretrained("Pulk17/Fake-News-Detection").to(device)
|
92 |
+
bert_model.eval()
|
93 |
+
|
94 |
+
# Detection function
|
95 |
+
def detect_fake_news(text):
|
96 |
+
inputs = bert_tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=512).to(device)
|
97 |
+
outputs = bert_model(**inputs)
|
98 |
+
probs = F.softmax(outputs.logits, dim=1)
|
99 |
+
predicted_class = torch.argmax(probs, dim=1).item()
|
100 |
+
confidence = torch.max(probs).item()
|
101 |
+
label = "FAKE π₯" if predicted_class == 0 else "REAL π©"
|
102 |
+
return f"Prediction: {label} ({confidence*100:.2f}% confidence)"
|
103 |
+
|
104 |
+
# Gradio interfaces
|
105 |
+
generator_interface = gr.Interface(
|
106 |
+
fn=gradio_fake_news_generator,
|
107 |
+
inputs=[
|
108 |
+
gr.Textbox(label="News Category (e.g., politics, tech)"),
|
109 |
+
gr.Textbox(label="Main Subject / Entity (e.g., Apple, Government)"),
|
110 |
+
gr.Textbox(label="Number of Headlines (1-10)")
|
111 |
+
],
|
112 |
+
outputs=gr.Textbox(label="Generated Fake News Headlines"),
|
113 |
+
title="π° Fake News Generator",
|
114 |
+
description="Generate fake news headlines using GPT-2"
|
115 |
+
)
|
116 |
+
|
117 |
+
detector_interface = gr.Interface(
|
118 |
+
fn=detect_fake_news,
|
119 |
+
inputs=gr.Textbox(lines=5, label="Enter full news text"),
|
120 |
+
outputs=gr.Textbox(label="Fake News Detection Result"),
|
121 |
+
title="π Fake News Detector",
|
122 |
+
description="Detect whether a news article is FAKE or REAL using BERT"
|
123 |
+
)
|
124 |
+
|
125 |
+
# Combine both interfaces
|
126 |
+
gr.TabbedInterface([generator_interface, detector_interface],
|
127 |
+
tab_names=["π° Generator", "π Detector"]).launch()
|
128 |
+
|