README.md

---
license: apache-2.0
tags:
  - fake-news-detection
  - multimodal
  - clip
  - bangla
  - english
  - supervised-learning
  - fact-checking
  - image-text
  - misinformation
  - fine-tuned
library_name: transformers
inference: true
language:
  - bn
  - en
---

# 🛡️ Multimodal BN-EN Fake News Scanner

A **fine-tuned CLIP model for detecting fake news in Bangla-English (BN-EN) content** using **text and image analysis**.

This model was **supervised-trained on real and fake news pairs** to better detect misinformation in South Asian digital content. During inference, it uses **prompt-based similarity** to classify inputs.

Deployed at:  
🔗 [https://huggingface.co/jarif/Multimodal-BNEN-Fake-News-Scanner-Model](https://huggingface.co/jarif/Multimodal-BNEN-Fake-News-Scanner-Model)

---

## ✅ Key Features

- ✅ **Supervised fine-tuning** on fake/real news dataset
- 📝 **Bangla + English text support**
- 🖼️ Analyzes **image authenticity**
- 🔍 Uses **prompt-based classification** with semantic similarity
- 🧩 Built on `openai/clip-vit-base-patch32`, fine-tuned for misinformation detection
- 🌍 Optimized for **South Asian context**
- 📦 Fully compatible with Hugging Face `transformers`

---

## 💻 Inference Code

Classify **images** and **text** as **Real** or **Fake** and display results in a clean table using `tabulate`.

### Install Dependencies

```python
pip install transformers torch pillow tabulate

from transformers import CLIPModel, CLIPProcessor
from PIL import Image
import torch
import torch.nn.functional as F
from tabulate import tabulate

# Load your fine-tuned model
model = CLIPModel.from_pretrained("jarif/Multimodal-BNEN-Fake-News-Scanner-Model")
processor = CLIPProcessor.from_pretrained("jarif/Multimodal-BNEN-Fake-News-Scanner-Model")

# Define class prompts in Bangla
class_texts = ["এটি ফেক নিউজ", "এটি রিয়েল নিউজ"]  # ["This is fake news", "This is real news"]

# --- Image Classification ---
image = Image.open("your_image.jpg").convert("RGB")  # Replace with your image path
image_inputs = processor(images=image, return_tensors="pt")
image_emb = model.get_image_features(**image_inputs)

# --- Text Classification ---
text = "পদ্মা নদীর প্রবল স্রোতে লঞ্চঘাট বিলীন হয়েছে।"
text_inputs = processor(text=text, return_tensors="pt", padding=True, truncation=True)
text_emb = model.get_text_features(**text_inputs)

# Get embeddings for class prompts
class_inputs = processor(text=class_texts, return_tensors="pt", padding=True, truncation=True)
class_embs = model.get_text_features(**class_inputs)

# Normalize embeddings (cosine similarity)
image_emb = F.normalize(image_emb, p=2, dim=-1)
text_emb = F.normalize(text_emb, p=2, dim=-1)
class_embs = F.normalize(class_embs, p=2, dim=-1)

# Compute similarity
image_sims = (image_emb @ class_embs.T).squeeze(0)
text_sims = (text_emb @ class_embs.T).squeeze(0)

# Predict
image_pred = image_sims.argmax().item()
text_pred = text_sims.argmax().item()

image_label = "🛑 Fake" if image_pred == 0 else "✅ Real"
text_label = "🛑 Fake" if text_pred == 0 else "✅ Real"

# Create result table
table = [
    ["ImageRelation", image_label],
    ["Text Relation", text_label]
]

# Print formatted table
print(tabulate(table, headers=["Modality", "Prediction"], tablefmt="fancy_grid"))
```


# 📊 Example Output
```
╒════════════════════════════════════════════════════════════════════╕
│                         Modality                │ Prediction │
╞════════════════════════════════════════════════════════════════════╡
│ Image Relation                                  │ ✅ Real    │
│ Text Relation                                   │ 🛑 Fake    │
╘════════════════════════════════════════════════════════════════════╛
```

# 📎 License
- This model is licensed under Apache 2.0.
- Fine-tuned by Sadik Al Jarif for public misinformation detection.

# 🙌 Acknowledgements
- OpenAI CLIP
- Hugging Face Transformers
- Bangla NLP Community